BEGIN:VCALENDAR PRODID;X-RICAL-TZSOURCE=TZINFO:-//com.denhaven2/NONSGML ri_cal gem//EN CALSCALE:GREGORIAN VERSION:2.0 METHOD:PUBLISH X-WR-TIMEZONE::Europe/Paris BEGIN:VTIMEZONE TZID;X-RICAL-TZSOURCE=TZINFO:Europe/Paris BEGIN:DAYLIGHT DTSTART:20090329T020000 RDATE:20090329T020000 TZOFFSETFROM:+0100 TZOFFSETTO:+0200 TZNAME:CEST END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T145000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T142500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4888 DESCRIPTION:This paper studies web object classification problem with the novel exploration of social tags. Automatically classifying web objects into manageable semantic categories has long been a fundamental preproc ess for indexing\, browsing\, searching\, and mining these objects. The explosive growth of heterogeneous web objects\, especially non-textual o bjects such as products\, pictures\, and videos\, has made the problem o f web classification increasingly challenging. Such objects often suffer from a lack of easy-extractable features with semantic information\, in terconnections between each other\, as well as training examples with ca tegory labels.\n\nIn this paper\, we explore the social tagging data to bridge this gap. We cast web object classification problem as an optimiz ation problem on a graph of objects and tags. We then propose an efficie nt algorithm which not only utilizes social tags as enriched semantic fe atures for the objects\, but also infers the categories of unlabeled obj ects from both homogeneous and heterogeneous labeled objects\, through t he implicit connection of social tags. Experiment results show that the exploration of social tags effectively boosts web object classification. Our algorithm significantly outperforms the state-of-the-art of general classification methods. SUMMARY:Exploring Social Tagging Graph for Web Object Classification LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4889 DESCRIPTION:User browsing information\, particularly their non-search rel ated activity\, reveals important contextual information on the preferen ces and the intent of web users. In this paper\, we expand the use of br owsing information for web search ranking and other applications\, with an emphasis on analyzing individual user sessions for creating aggregate models. In this context\, we introduce ClickRank\, an efficient\, scala ble algorithm for estimating web page and web site importance from brows ing information. We lay out the theoretical foundation of ClickRank base d on an intentional surfer model and analyze its properties. We evaluate its effectiveness for the problem of web search ranking\, showing that it contributes significantly to retrieval performance as a novel web sea rch feature. We demonstrate that the results produced by ClickRank for w eb search ranking are highly competitive with those produced by other ap proaches\, yet achieved at better scalability and substantially lower co mputational costs. Finally\, we discuss novel applications of ClickRank in providing enriched user web search experience\, highlighting the usef ulness of our approach for non-ranking tasks. SUMMARY:Mining Rich Session Context to Improve Web Search LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T120000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T114500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4890 DESCRIPTION:Co-clustering is based on the duality between data points (e. g. documents) and features (e.g. words)\, i.e. data points can be groupe d based on their distribution on features\, while features can be groupe d based on their distribution on the data points. In the past decade\, s everal co-clustering algorithms have been proposed and shown to be super ior to traditional one-side clustering. However\, existing co-clustering algorithms fail to consider the geometric structure in the data\, which is essential for clustering data on manifold. To address this problem\, in this paper\, we propose a Dual Regularized Co-Clustering (DRCC) meth od based on semi-nonnegative matrix tri-factorization. We deem that not only the data points\, but also the features are sampled from some manif olds\, namely data manifold and feature manifold respectively. As a resu lt\, we construct two graphs\, i.e. data graph and feature graph\, to ex plore the geometric structure of data manifold and feature manifold. The n our co-clustering method is formulated as semi-nonnegative matrix tri- factorization with two graph regularizers\, requiring that the cluster l abels of data points are smooth with respect to the data manifold\, whil e the cluster labels of features are smooth with respect to the feature manifold. We will show that DRCC can be solved via alternating minimizat ion\, and its convergence is theoretically guaranteed. Experiments of cl ustering on many benchmark data sets demonstrate that the proposed metho d outperforms many state of the art clustering methods. SUMMARY:Co-Clustering on Manifolds LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T170500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T165000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4891 DESCRIPTION:The central challenge in temporal data analysis is to obtain knowledge about its underlying dynamics. In this paper\, we address the observation of noisy\, stochastic processes and attempt to detect tempo ral segments that are related to inconsistencies and irregularities in i ts dynamics. Many conventional anomaly detection approaches detect anoma lies based on the distance between patterns\, and often provide only lim ited intuition about the generative process of the anomalies. Meanwhile \, model-based approaches have difficulty in identifying a small\, clust ered set of anomalies.\n\nWe propose Information-theoretic Meta-clusteri ng (ITMC)\, a formalization of model-based clustering principled by the theory of lossy data compression. ITMC identifies a `unique' cluster who se distribution diverges significantly from the entire dataset. Furtherm ore\, ITMC employs a regularization term derived from the preference for high compression rate\, which is critical to the precision of detection .\n\nFor empirical evaluation\, we apply ITMC to two temporal anomaly de tection tasks. Datasets are taken from generative processes involving he terogeneous and inconsistent dynamics. A comparison to baseline methods shows that the proposed algorithm detects segments from irregular states with significantly high precision and recall. SUMMARY:Detection of Unique Temporal Segments by Information Theoretic Me ta-clustering LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T170500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T165000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4892 DESCRIPTION:We present a new approach to large-scale graph mining based o n so-called backbone refinement classes. The method efficiently mines tr ee-shaped subgraph descriptors under minimum frequency and significance constraints\, using classes of fragments to reduce feature set size and running times.\n\nThe classes are defined in terms of fragments sharing a common backbone. The method is able to optimize structural inter-featu re entropy as opposed to occurrences\, which is characteristic for open or closed fragment mining. In the experiments\, the proposed method redu ces feature set sizes by >90 % and >30 % compared to complete tree minin g and open tree mining\, respectively. Evaluation using crossvalidation runs shows that their classification accuracy is similar to the complete set of trees but significantly better than that of open trees. Compared to open or closed fragment mining\, a large part of the search space ca n be pruned due to an improved statistical constraint (dynamic upper bou nd adjustment)\, which is also confirmed in the experiments in lower run ning times compared to ordinary (static) upper bound pruning. Further an alysis using large-scale datasets yields insight into important properti es of the proposed descriptors\, such as the dataset coverage and the cl ass size represented by each descriptor. A final cross-validation run co nfirms that the novel descriptors render large training sets feasible wh ich previously might have been intractable. SUMMARY:Large-Scale Graph Mining Using Backbone Refinement Classes LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4893 DESCRIPTION:We study in this paper the problem of incremental crawling of web forums\, which is a very fundamental yet challenging step in many w eb applications. Traditional approaches mainly focus on scheduling the r evisiting strategy of each individual page. However\, simply assigning d ifferent weights for different individual pages is usually inefficient i n crawling forum sites because of the different characteristics between forum sites and general websites. Instead of treating each individual pa ge independently\, we propose a list-wise strategy by taking into accoun t the site-level knowledge. Such site-level knowledge is mined through r econstructing the linking structure\, called sitemap\, for a given forum site. With the sitemap\, posts from the same thread but distributed on various pages can be concatenated according to their timestamps. After t hat\, for each thread\, we employ a regression model to predict the time when the next post arrives. Based on this model\, we develop an efficie nt crawler which is 260% faster than some state-of-the-art methods in te rms of fetching new generated content\; and meanwhile our crawler also e nsure a high coverage ratio. Experimental results show promising perform ance of Coverage\, Bandwidth utilization\, and Timeliness of our crawler on 18 various forums. SUMMARY:Incorporating Site-Level Knowledge for Incremental Crawling of We b Forums: A List-wise Strategy LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T120000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T114500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4894 DESCRIPTION:This paper studies the problem of frequent pattern mining wi th uncertain data. We will show how broad classes of algorithms can be e xtended to the uncertain data setting. In particular\, we will study can didate generate-and-test algorithms\, hyper-structure algorithms and p attern growth based algorithms. One of our insightful observations is t hat the experimental behavior of different classes of algorithms is very different in the uncertain case as compared to the deterministic case. In particular\, the hyper-structure and the candidate generate-and-test algorithms perform much better than tree-based algorithms. This counter -intuitive behavior is an important observation from the perspective of algorithm design of the uncertain variation of the problem. We will tes t the approach on a number of real and synthetic data sets\, and show th e effectiveness of two of our approaches over competitive techniques.\n\ nExecutable and Data Sets: Available at: http://dbgroup.cs.tsinghua.edu .cn/liyan/u_mining.tar.gz SUMMARY:Frequent Pattern Mining with Uncertain Data LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T114500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4895 DESCRIPTION:Rock art is an archaeological term for human-made markings on stone. It is believed that there are millions of petroglyphs in North A merica alone\, and the study of this valued cultural resource has implic ations even beyond anthropology and history. Surprisingly\, although ima ge processing\, information retrieval and data mining have had large imp acts on many human endeavors\, they have had essentially zero impact on the study of rock art. In this work we identify the reasons for this\, a nd introduce a novel distance measure and algorithms which allow efficie nt and effective data mining of large collections of rock art. SUMMARY:Augmenting the Generalized Hough Transform to Enable the Mining o f Petroglyphs LOCATION:La Seine C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4896 DESCRIPTION:Classification of time series has been attracting great inter est over the past decade. Recent empirical evidence has strongly suggest ed that the simple nearest neighbor algorithm is very difficult to beat for most time series problems. While this may be considered good news\, given the simplicity of implementing the nearest neighbor algorithm\, th ere are some negative consequences of this. First\, the nearest neighbor algorithm requires storing and searching the entire dataset\, resulting in a time and space complexity that limits its applicability\, especial ly on resource-limited sensors. Second\, beyond mere classification accu racy\, we often wish to gain some insight into the data. \nIn this work we introduce a new time series primitive\, time series shapelets\, which addresses these limitations. Informally\, shapelets are time series sub sequences which are in some sense maximally representative of a class. A s we shall show with extensive empirical evaluations in diverse domains\ , algorithms based on the time series shapelet primitives can be interpr etable\, more accurate and significantly faster than state-of-the-art cl assifiers. SUMMARY:Time Series Shapelets: A New Primitive for Data Mining LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4897 DESCRIPTION:Spatial classification is the task of learning models to pred ict class labels based on the features of entities as well as the spatia l relationships to other entities and their features. Spatial data can b e represented as multi-relational data\, however it presents novel chall enges not present in multi-relational problems. One such problem is that spatial relationships are embedded in space\, unknown a priori\, and it is part of the algorithms task to determine which relationships are im portant and what properties to consider. In order to determine when two entities are spatially related in an adaptive and non-parametric way\, w e propose a Voronoi-based neighbourhood definition upon which spatial li terals can be built. Properties of these neighbourhoods also need to be described and used for classification purposes. Non-spatial aggregation literals already exist within the multi-relational framework\, but are n ot sufficient for comprehensive spatial classification. A formal set of additions to the multi-relational data mining framework is proposed\, to be able to represent spatial aggregations as well as spatial features a nd literals. These additions allow for capturing more complex interactio ns and spatial occurrences such as spatial trends. In order to more effi ciently perform the rule learning and exploit powerful multi-processor m achines\, a scalable parallelized method capable of reducing the runtime by several factors is presented. The method is compared against existin g methods by experimental evaluation on a real world crime dataset which demonstrate the importance of the neighbourhood definition and the adva ntages of parallelization. SUMMARY:A Multi-Relational Approach to Spatial Classification LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T172000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T170500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4898 DESCRIPTION:Counting the number of triangles in a graph is a beautiful al gorithmic problem which has gained importance over the last years due to its significant role in complex network analysis. Metrics frequently computed such as the clustering coefficient and the transitivity ratio involve the execution of a triangle counting algorithm. Furthermore\, s everal interesting graph mining applications rely on computing the numbe r of triangles in the graph of interest. \n\nIn this paper\, we focus on the problem of counting triangles in a graph. We propose a practical me thod\, out of which all triangle counting algorithms can potentially ben efit. Using a straight-forward triangle counting algorithm as a black bo x\, we performed 166 experiments on real-world networks and on synthetic datasets as well\, where we show that our method works with high accur acy\, typically more than 99\\% and gives significant speedups\, resulti ng in even $\\approx$ 130 times faster performance. SUMMARY:Doulion: Counting triangles in massive graphs with a coin LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T123500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T122000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4899 DESCRIPTION:This paper addresses Named Entity Mining (NEM)\, in which we mine knowledge about named entities such as movies\, games\, and books f rom a huge amount of data. NEM is potentially useful in many application s including web search\, online advertisement\, and recommender system. There are three challenges for the task: finding suitable data source\, coping with the ambiguities of named entity classes\, and incorporating necessary human supervision into the mining process. This paper proposes conducting NEM by using click-through data collected at a web search en gine\, employing a topic model that generates the click-through data\, a nd learning the topic model by weak supervision from humans. Specificall y\, it characterizes each named entity by its associated queries and URL s in the click-through data. It uses the topic model to resolve ambiguit ies of named entity classes by representing the classes as topics. It em ploys a method\, referred to as Weakly Supervised Latent Dirichlet Alloc ation (WS-LDA)\, to accurately learn the topic model with partially labe led named entities. Experiments on a large scale click-through data cont aining over 1.5 billion query-URL pairs show that the proposed approach can conduct very accurate NEM and significantly outperforms the baseline . SUMMARY:Named Entity Mining from Click-Through Data Using Weakly Supervis ed Latent Dirichlet Allocation LOCATION:Louis Armstrong A\,B\,C+D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4900 DESCRIPTION:BEST RESEARCH PAPER AWARD WINNER\n\nCustomer preferences for products are drifting over time. Product perception and popularity are c onstantly changing as new selection emerges. Similarly\, customer inclin ations are evolving\, leading them to ever redefine their taste. Thus\, modeling temporal dynamics should be a key when designing recommender sy stems or general customer preference models. However\, this raises uniqu e challenges. Within the eco-system intersecting multiple products and c ustomers\, many different characteristics are shifting simultaneously\, while many of them influence each other and often those shifts are delic ate and associated with a few data instances. This distinguishes the pro blem from concept drift explorations\, where mostly a single concept is tracked. Classical time-window or instance-decay approaches cannot work\ , as they lose too much signal when discarding data instances. A more s ensitive approach is required\, which can make better distinctions betw een transient effects and long term patterns. The paradigm we offer is c reating a model tracking the time changing behavior throughout the life span of the data. This allows us to exploit the relevant components of a ll data instances\, while discarding only what is modeled as being irrel evant. Accordingly\, we revamp two leading collaborative filtering recom mendation approaches. Evaluation is made on a large movie rating dataset by Netflix. Results are encouraging and better than those previously re ported on this dataset. SUMMARY:Collaborative Filtering with Temporal Dynamics LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T122000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T120500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4901 DESCRIPTION:Automatic news extraction from news pages is important in man y Web applications such as news aggregation. However\, the existing news extraction methods based on template-level wrapper induction have three serious limitations. First\, the existing methods cannot correctly extr act pages belonging to an unseen template. Second\, it is costly to main tain up-to-date wrappers for a large amount of news websites\, because a ny change of a template may invalidate the corresponding wrapper. Last\, the existing methods can merely extract unformatted plain texts\, and t hus are not user friendly. In this paper\, we tackle the problem of temp late-independent Web news extraction in a user-friendly way. We formaliz e Web news extraction as a machine learning problem and learn a template -independent wrapper using a very small number of labeled news pages fro m a single site. Novel features dedicated to news titles and bodies are developed. Correlations between news titles and news bodies are exploite d. Our template-independent wrapper can extract news pages from differen t sites regardless of templates. Moreover\, our approach can extract not only texts\, but also images and animates within the news bodies and th e extracted news articles are in the same visual style as in the origina l pages. In our experiments\, a wrapper learned from 40 pages from a sin gle news site achieved an accuracy of 98.1% on 3\,973 news pages from 12 news sites. SUMMARY:Can We Learn a Template-Independent Wrapper for News Article Extr action from a Single Training Site? LOCATION:Louis Armstrong A\,B\,C+D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4902 DESCRIPTION:Given a spatial data set placed on an n × n grid\, our goal is to find the rectangular regions within which subsets of the data set exhibit anomalous behavior. We develop algorithms that\, given any user- supplied arbitrary likelihood function\, conduct a likelihood ratio hypo thesis test (LRT) over each rectangular region in the grid\, rank all of the rectangles based on the computed LRT statistics\, and return the to p few most interesting rectangles. To speed this process\, we develop me thods to prune rectangles without computing their associated LRT statist ics. SUMMARY:A LRT Framework for Fast Spatial Anomaly Detection LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4903 DESCRIPTION:Clustering validation is a long standing challenge in the clu stering literature. While many validation measures have been developed f or evaluating the performance of clustering algorithms\, these measures often provide inconsistent information about the clustering performance and the best suitable measures to use in practice remain unknown. This p aper thus fills this crucial void by giving an organized study of 16 ext ernal validation measures for K-means clustering. Specifically\, we firs t introduce the importance of measure normalization in the evaluation of the clustering performance on data with imbalanced class distributions. We also provide normalization solutions for several measures. In additi on\, we summarize the major properties of these external measures. These properties can serve as the guidance for the selection of validation me asures in different application scenarios. Finally\, we reveal the inter relationships among these external measures. By mathematical transformat ion\, we show that some validation measures are equivalent. Also\, some measures have consistent validation performances. Most importantly\, we provide a guide line to select the most suitable validation measures for K-means clustering. SUMMARY:Adapting the Right Measures for K-means Clustering LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T121500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4904 DESCRIPTION:Data analytics tools and frameworks abound\, yet rapid deploy ment of analytics solutions that deliver actionable insights from busine ss data remains a challenge. The primary reason is that on-field practit ioners are required to be both technically proficient and knowledgeable about the business. \n\nThe recent abundance of unstructured business da ta has thrown up new opportunities for analytics\, but has also multipli ed the deployment challenge\, since interpretation of concepts derived f rom textual sources require a deep understanding of the business. In su ch a scenario\, a managed service for analytics comes up as the best alt ernative. A managed analytics service is centered around a business anal yst who acts as a liaison between the business and the technology. This calls for new tools that assist the analyst to be {\\it efficient} in t he tasks that she needs to execute. Also\, the analytics needs to be {\ \it repeatable}\, in that the delivered insights should not depend heavi ly on the expertise of specific analysts. These factors lead us to iden tify new areas that open up for KDD research in terms of `time-to-insigh t' and repeatability for these analysts. We present our analytics frame work in the form of a managed service offering for CRM analytics. We des cribe different analyst-centric tools using a case study from real-life engagements and demonstrate their effectiveness. SUMMARY:Enabling Analysts in Managed Services for CRM Analytics LOCATION:La Seine C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T120000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T114500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4905 DESCRIPTION:Identifying similar keywords\, known as broad matches\, is an important task in online advertising that has become a standard feature on all major keyword advertising platforms. Effective broad matching le ads to improvements in both relevance and monetization\, while increasin g advertisers' reach and making campaign management easier. In this pape r\, we present a learning-based approach to broad matching that is based on exploiting implicit feedback in the form of advertisement clickthrou gh logs. Our method can utilize arbitrary similarity functions by incorp orating them as features. We present an online learning algorithm\, Amne siac Averaged Perceptron\, that is highly efficient yet able to quickly adjust to the rapidly-changing distributions of bidded keywords\, advert isements and user behavior. Experimental results obtained from (1) histo rical logs and (2) live trials on a large-scale advertising platform dem onstrate the effectiveness of the proposed algorithm and the overall suc cess of our approach in identifying high-quality broad match mappings. SUMMARY:Catching the Drift: Learning Broad Matches from Clickthrough Data LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4906 DESCRIPTION:There is a wide variety of data mining methods available\, an d it is generally useful in exploratory data analysis to use many differ ent methods for the same dataset. This\, however\, leads to the problem of whether the results found by one method are a reflection of the pheno menon shown by the results of another method\, or whether the results de pict in some sense unrelated properties of the data. For example\, using clustering can give indication of a clear cluster structure\, and compu ting correlations between variables can show that there are many signifi cant correlations in the data. However\, it can be the case that the cor relations are actually determined by the cluster structure.\n\nIn this p aper\, we consider the problem of randomizing data so that previously di scovered patterns or models are taken into account. The randomization me thods can be used in iterative data mining. At each step in the data min ing process\, the randomization produces random samples from the set of data matrices satisfying the already discovered patterns or models. That is\, given a data set and some statistics (e.g.\, cluster centers or co -occurrence counts) of the data\, the randomization methods sample data sets having similar values of the given statistics as the original data set. We use Metropolis sampling based on local swaps to achieve this. We describe experiments on real data that demonstrate the usefulness of ou r approach. Our results indicate that in many cases\, the results of\, e .g.\, clustering actually imply the results of\, say\, frequent pattern discovery. SUMMARY:Tell Me Something I Don't Know: Randomization Strategies for Iter ative Data Mining LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T120500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T115000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4907 DESCRIPTION:Web content analysis often has two sequential and separate st eps: Web Classification to identify the target Web pages\, and Web Infor mation Extraction to extract the metadata contained in the target Web pa ges. This decoupled strategy is highly ineffective since the errors in W eb classification will be propagated to Web information extraction and e ventually accumulate to a high level. In this paper we study the mutual dependencies between these two steps and propose to combine them by usin g a model of Conditional Random Fields (CRFs). This model can be used to simultaneously recognize the target Web pages and extract the correspon ding metadata. Systematic experiments in our project OfCourse for online course search show that this model significantly improves the F1 value for both of the two steps. We believe that our model can be easily gener alized to many Web applications. SUMMARY:Towards Combining Web Classification and Web Information Extracti on: A Case Study LOCATION:Louis Armstrong A\,B\,C+D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T153000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T151500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4908 DESCRIPTION:When labeled examples are limited and difficult to obtain\, t ransfer learning employs knowledge from a source domain to improve learn ing accuracy in the target domain. However\, the assumption made by exis ting approaches\, that the marginal and conditional probabilities are di rectly related between source and target domains\, has limited applicabi lity in either the original space or its linear\ntransformations. To sol ve this problem\, we propose an adaptive kernel approach that maps the m arginal distribution of target-domain and source-domain data into a comm on kernel space\, and utilize a sample selection strategy to draw condit ional probabilities between the two domains closer. We formally show tha t under the kernel-mapping space\, the difference in distributions betwe en the\ntwo domains is bounded\; and the prediction error of the propose d approach can also be bounded. Experimental results demonstrate that th e proposed method outperforms both traditional inductive classifiers and the state-of-the-art boosting-based transfer algorithms on most domains \, including text categorization and web page ratings. In particular\, i t can achieve around 10% higher accuracy than other approaches for the t ext categorization problem. The source code and datasets are available f rom the authors. SUMMARY:Cross Domain Distribution Adaptation via Kernel Mapping LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T113500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4909 DESCRIPTION:Learning probabilistic graphical models from high-dimensional datasets is a computationally challenging task. In many interesting app lications\, the domain dimensionality is such as to prevent state-of-the -art statistical learning techniques from delivering accurate models in reasonable time. This paper presents a hybrid random field model for pse udo-likelihood estimation in high-dimensional domains. A theoretical ana lysis proves that the class of pseudo-likelihood distributions represent able by hybrid random fields strictly includes the class of joint probab ility distributions representable by Bayesian networks. In order to lear n hybrid random fields from data\, we develop the Markov Blanket Merging algorithm. Theoretical and experimental evidence shows that Markov Blan ket Merging scales up very well to high-dimensional datasets. As compare d to other widely used statistical learning techniques\, Markov Blanket Merging delivers accurate results in a number of link prediction tasks\, while achieving also significant improvements in terms of computational efficiency.\n\nOur software implementation of the models investigated i n this paper is publicly available at http://www.dii.unisi.it/~freno/. T he same website also hosts the datasets used in this work that are not a vailable elsewhere in the same preprocessing used for our experiments. SUMMARY:Scalable Pseudo-Likelihood Estimation in Hybrid Random Fields LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T115000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T113500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4910 DESCRIPTION:Nowadays\, enormous amounts of data are continuously generate d not only in massive scale\, but also from different\, sometimes confli cting\, views. Therefore\, it is important to consolidate different conc epts for intelligent decision making. For example\, to predict the resea rch areas of some people\, the best results are usually achieved by comb ining and consolidating predictions obtained from the publication networ k\, co-authorship network and the textual content of their publications. Multiple supervised and unsupervised hypotheses can be drawn from these information sources\, and negotiating their differences and consolidati ng decisions usually yields a much more accurate model due to the divers ity and heterogeneity of these models. In this paper\, we address the pr oblem of "consensus learning" among competing hypotheses\, which either rely on outside knowledge (supervised learning) or internal structure (u nsupervised clustering). We argue that consensus learning is an NP-hard problem and thus propose to solve it by an efficient heuristic method. W e construct a belief graph to first propagate predictions from supervise d models to the unsupervised\, and then negotiate and reach consensus am ong them. Their final decision is further consolidated by calculating ea ch model's weight based on its degree of consistency with other models. Experiments are conducted on 20 Newsgroups data\, Cora research papers\, DBLP author-conference network\, and Yahoo! Movies datasets\, and the r esults show that the proposed method improves the classification accurac y and the clustering quality measure (NMI) over the best base model by u p to 10%. Furthermore\, it runs in time proportional to the number of in stances\, which is very efficient for large scale data sets. SUMMARY:Heterogeneous Source Consensus Learning via Decision Propagation and Negotiation LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T142500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4911 DESCRIPTION:In several applications involving regression or classificatio n\, along with making predictions it is important to assess how accurate or reliable individual predictions are. This is particularly important in cases where due to finite resources or domain requirements\, one want s to make decisions based only on the most reliable rather than on the e ntire set of predictions. This paper introduces novel and effective ways of ranking predictions by their accuracy for problems involving large-s cale\, heterogeneous data with a dyadic structure\, i.e.\, where the ind ependent variables can be naturally decomposed into three groups associa ted with two sets of elements and their combination. These approaches ar e based on modeling the data by a collection of localized models learnt while simultaneously partitioning (co-clustering) the data. For regress ion this leads to the concept of "certainty lift". We also develop a ro bust predictive modeling technique that identifies and models only the m ost coherent regions of the data to give high predictive accuracy on the selected subset of response values. Extensive experimentation on real l ife datasets highlights the utility of our proposed approaches. SUMMARY:Mining for the Most Certain Predictions from Dyadic Data LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T153000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T151500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4912 DESCRIPTION:The increasing availability of electronic communication data\ , such as that arising from e-mail exchange\, presents social and inform ation scientists with new possibilities for characterizing individual be havior and\, by extension\, identifying latent structure in human popula tions. Here\, we propose a model of individual e-mail communication tha t is sufficiently rich to capture meaningful variability across individu als\, while remaining simple enough to be interpretable. We show that t he model\, a cascading non-homogeneous Poisson process\, can be formulat ed as a double-chain hidden Markov model\, allowing us to use an efficie nt inference algorithm to estimate the model parameters from observed da ta. We then apply this model to two e-mail data sets consisting of 404 a nd 6\,164 users\, respectively\, that were collected from two universiti es in different countries and years. We find that the resulting best-es timate parameter distributions for both data sets are surprisingly simil ar\, indicating that at least some features of communication dynamics ge neralize beyond specific contexts. We also find that variability of ind ividual behavior over time is significantly less than variability across the population\, suggesting that individuals can be classified into per sistent "types". We conclude that communication patterns may prove usef ul as an additional class of attribute data\, complementing demographic and network data\, for user classification and outlier detection---a poi nt that we illustrate with an interpretable clustering of users based on their inferred model parameters. SUMMARY:Characterizing Individual Communication Patterns LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T170500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T165000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4913 DESCRIPTION:Many applications in surveillance\, monitoring\, scientific d iscovery\, and data cleaning require the identification of anomalies. Al though many methods have been developed to identify statistically signif icant anomalies\, a more difficult task is to identify anomalies that ar e both interesting and statistically significant. Category detection is an emerging area of machine learning that can help address this issue us ing a "human-in-the-loop" approach. In this interactive setting\, the al gorithm asks the user to label a query data point under an existing cate gory or declare the query data point to belong to a previously undiscove red category. The goal of category detection is to bring to the user's a ttention a representative data point from each category in the data in a s few queries as possible. In a data set with imbalanced categories\, th e main challenge is in identifying the rare categories or anomalies\; he nce\, the task is often referred to as {\\it rare} category detection. W e present a new approach to rare category detection based on hierarchica l mean shift. In our approach\, a hierarchy is created by repeatedly app lying mean shift with an increasing bandwidth on the data. This hierarch y allows us to identify anomalies in the data set at different scales\, which are then posed as queries to the user. The main advantage of this methodology over existing approaches is that it does not require any kno wledge of the dataset properties such as the total number of categories or the prior probabilities of the categories. Results on real-world data sets show that our hierarchical mean shift approach performs consistent ly better than previous techniques. SUMMARY:Category Detection Using Hierarchical Mean Shift LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T150500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T145000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4914 DESCRIPTION:Controlled experiments\, also called randomized experiments a nd A/B tests\, have had a profound influence on multiple fields\, includ ing medicine\, agriculture\, manufacturing\, and advertising. While the theoretical aspects of offline controlled experiments have been well stu died and documented\, the practical aspects of running them in online se ttings\, such as web sites and services\, are still being developed. As the usage of controlled experiments grows in these online settings\, it is becoming more important to understand the opportunities and pitfalls one might face when using them in practice. A survey of online controlle d experiments and lessons learned were previously documented in Controll ed Experiments on the Web: Survey and Practical Guide (Kohavi\, et al.\, 2009). In this follow-on paper\, we focus on pitfalls we have seen afte r running numerous experiments at Microsoft. The pitfalls include a wid e range of topics\, such as assuming that common statistical formulas us ed to calculate standard deviation and statistical power can be applied and ignoring robots in analysis (a problem unique to online settings). O nline experiments allow for techniques like gradual ramp-up of treatment s to avoid the possibility of exposing many customers to a bad (e.g.\, b uggy) Treatment. With that ability\, we discovered that its easy to inc orrectly identify the winning Treatment because of Simpsons paradox. SUMMARY:Seven Pitfalls to Avoid when Running Controlled Experiments on th e Web LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T115000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T112500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4915 DESCRIPTION:Learning from data streams is a research area of increasing i mportance. Nowadays\, several stream learning algorithms have been devel oped. Most of them learn decision models that continuously evolve over t ime\, run in resource-aware environments\, detect and react to changes i n the environment generating data. One important issue\, not yet conveni ently addressed\, is the design of experimental work to evaluate and com pare decision models that evolve over time. There are no golden standar ds for assessing performance in non-stationary environments. This paper proposes a general framework for assessing predictive stream learning al gorithms. We defend the use of Predictive Sequential methods for error e stimate -- the prequential error. The prequential error allows us to mon itor the evolution of the performance of models that evolve over time. N evertheless\, it is known to be a pessimistic estimator in comparison to holdout estimates. To obtain more reliable estimators we need some forg etting mechanism. Two viable alternatives are: sliding windows and fadin g factors. We observe that the prequential error converges to an holdout estimator when estimated over a sliding window or using fading factors. %A similar observation applies for fading factors. We present illustrat ive examples of the use of prequential error estimators\, using fading f actors\, for the tasks of: \n i. assessing performance of a learning alg orithm\; \n ii. comparing learning algorithms\;\n iii. hypothesis testin g using McNemar test\; and \n iv. change detection using Page-Hinkley te st.\n\nIn these tasks\, the prequential error estimated using fading fac tors provide reliable estimators. In comparison to sliding windows\, fa ding factors are faster and memory-less\, a requirement for streaming ap plications. This paper is a contribution to a discussion in the good-pra ctices on performance assessment when learning dynamic models that evolv e over time. SUMMARY:Issues in Evaluation of Stream Learning Algorithms LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T171500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T165000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4916 DESCRIPTION:Tracking new topics\, ideas\, and "memes" across the Web has been an issue of considerable interest. Recent work has developed method s for tracking topic shifts over long time scales\, as well as abrupt sp ikes in the appearance of particular named entities. However\, these app roaches are less well suited to the identification of content that sprea ds widely and then fades over time scales on the order of days --- the t ime scale at which we perceive news and events.\n\nWe develop a framewor k for tracking short\, distinctive phrases that travel relatively intact through on-line text\; developing scalable algorithms for clustering te xtual variants of such phrases\, we identify a broad class of memes that exhibit wide spread and rich variation on a daily basis. As our princip al domain of study\, we show how such a meme-tracking approach can provi de a coherent representation of the news cycle --- the daily rhythms in the news media that have long been the subject of qualitative interpreta tion but have never been captured accurately enough to permit actual qua ntitative analysis. We tracked 1.6 million mainstream media sites and bl ogs over a period of three months with the total of 90 million articles and we find a set of novel and persistent temporal patterns in the news cycle. In particular\, we observe a typical lag of 2.5 hours between the peaks of attention to a phrase in the news media and in blogs respectiv ely\, with divergent behavior around the overall peak and a ``heartbeat' '-like pattern in the handoff between news and blogs. We also develop an d analyze a mathematical model for the kinds of temporal variation that the system exhibits. SUMMARY:Meme-tracking and the dynamics of the news cycle LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T142500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4917 DESCRIPTION:Overall performance of the data mining process depends not ju st on the value of the induced knowledge but also on various costs of th e process itself such as the cost of acquiring and pre-processing traini ng examples\, the CPU cost of model induction\, and the cost of committe d errors. Recently\, several progressive sampling strategies for maximiz ing the overall data mining utility have been proposed. All these strate gies are based on repeated acquisitions of additional training examples until a utility decrease is observed. In this paper\, we present an alte rnative\, projective sampling strategy\, which fits functions to a parti al learning curve and a partial run-time curve obtained from a small sub set of potentially available data and then uses these projected function s to analytically estimate the optimal training set size. The proposed a pproach is evaluated on a variety of benchmark datasets using the RapidM iner environment for machine learning and data mining processes. The res ults show that the learning and run-time curves projected from only seve ral data points can lead to a cheaper data mining process than the commo n progressive sampling methods. SUMMARY:Improving Data Mining Utility with Projective Sampling LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T121500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T120000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4918 DESCRIPTION:Spectral clustering refers to a flexible class of clustering procedures that can produce high-quality clusterings on small data sets but which has limited applicability to large-scale problems due to its c omputational complexity of O(n3) in general\, with n the number of data points. We extend the range of spectral clustering by developing a gener al framework for fast approximate spectral clustering in which a distort ion-minimizing local transformation is first applied to the data. This f ramework is based on a theoretical analysis that provides a statistical characterization of the effect of local distortion on the mis-clustering rate. We develop two concrete instances of our general framework\, one based on local k-means clustering (KASP) and one based on random project ion trees (RASP). Extensive experiments show that these algorithms can achieve significant speedups with little degradation in clustering accur acy. Specifically\, our algorithms outperform k-means by a large margin in terms of accuracy\, and run several times faster than approximate spe ctral clustering based on the Nystrom method\, with comparable accuracy and significantly smaller memory footprint. Remarkably\, our algorithms make it possible for a single machine to spectral cluster data sets with a million observations within several minutes. SUMMARY:Fast Approximate Spectral Clustering LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T171500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T165000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4919 DESCRIPTION:Best Student Paper Award Winner\n\nSharing healthcare data ha s become a vital requirement in healthcare system management\; however\, inappropriate sharing and usage of healthcare data could threaten patie nts' privacy. In this paper\, we study the privacy concerns of the blood transfusion information-sharing system between the Hong Kong Red Cross Blood Transfusion Service (BTS) and public hospitals\, and identify the major challenges that make traditional data anonymization methods not ap plicable. Furthermore\, we propose a new privacy model called LKC-privac y\, together with an anonymization algorithm\, to meet the privacy and i nformation requirements in this BTS case. Experiments on the real-life d ata demonstrate that our anonymization algorithm can effectively retain the essential information in anonymous data for data analysis and is sca lable for anonymizing large datasets. SUMMARY:Anonymizing Healthcare Data: A Case Study on the Blood Transfusio n Service LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T170500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T165000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4920 DESCRIPTION:The problem of ethnicity identification from names has a vari ety of important applications\, including biomedical research\, demograp hic studies\, and marketing. Here we report on the development of an et hnicity classifier where all training data is extracted from public\, no n-confidential (and hence somewhat unreliable) sources. Our classifier uses hidden Markov models (HMMs) and decision trees to classify names in to 13 cultural/ethnic groups with individual group accuracy comparable a ccuracy to earlier binary (e.g.\, Spanish/non-Spanish) classifiers. We have applied this classifier to over 20 million names from a large-scale news corpus\, identifying interesting temporal and spatial trends on th e representation of particular cultural/ethnic groups. SUMMARY:Name-Ethnicity Classification from Open Sources LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T124500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T123000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4921 DESCRIPTION:In realistic settings the prevalence of a class may change af ter a classifier is induced and this will degrade the performance of the classifier. Further complicating this scenario is the fact that labeled data is often scarce and expensive. In this paper we address the proble m where the class distribution changes and only unlabeled examples are a vailable from the new distribution. We design and evaluate a number of m ethods for coping with this problem and compare the performance of these methods. Our quantification-based methods estimate the class distributi on of the unlabeled data from the changed distribution and adjust the or iginal classifier accordingly\, while our semi-supervised methods build a new classifier using the examples from the new (unlabeled) distributio n which are supplemented with predicted class values. We also introduce a hybrid method that utilizes both quantification and semi-supervised le arning. All methods are evaluated using accuracy and F-measure on a set of benchmark data sets. Our results demonstrate that our methods yield s ubstantial improvements in accuracy and F-measure. SUMMARY:Quantification and Semi-supervised Classification Methods for Han dling Changes in Class Distribution LOCATION:La Seine C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T144000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T142500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4922 DESCRIPTION:Influence maximization is the problem of finding a small subs et of nodes (seed nodes) in a social network that could maximize the spr ead of influence. In this paper\, we study the efficient influence maxi mization from two complementary directions. One is to improve the origi nal greedy algorithm of [KKT03] and its improvement [LKGFVG07] to furthe r reduce its running time\, and the second is to propose new degree disc ount heuristics that improves influence spread. We evaluate our algorit hms by experiments on two large academic collaboration graphs obtained f rom the online archival database arXiv.org. Our experimental results sh ow that (a) our improved greedy algorithm achieves better running time c omparing with the improvement of [LKGFVG07] with matching influence spre ad\, (b) our degree discount heuristics achieve much better influence sp read than classic degree and centrality-based heuristics\, and when tune d for a specific influence cascade model\, it achieves almost matching i nfluence thread with the greedy algorithm\, and more importantly (c) the degree discount heuristics run only in milliseconds while even the impr oved greedy algorithms run in hours in our experiment graphs with a few tens of thousands of nodes.\n\nBased on our results\, we believe that fi ne-tuned heuristics may provide truly scalable solutions to the influenc e maximization problem with satisfying influence spread and blazingly fa st running time. Therefore\, contrary to what implied by the conclusion of [KKT03] that traditional heuristics are outperformed by the greedy ap proximation algorithm\, our results shed new lights on the research of h euristic algorithms. SUMMARY:Efficient Influence Maximization in Social Networks LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T120500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T115000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4923 DESCRIPTION:The pervasiveness of mobile devices and location based servic es is leading to an increasing volume of mobility data. This side effect provides the opportunity for innovative methods that analyze the behavi ors of movements.\n\nIn this paper we propose WhereNext\, which is a met hod aimed at predicting with a certain level of accuracy the next locati on of a moving object. The prediction uses previously extracted movement patterns named Trajectory Patterns\, which are a concise representation of behaviors of moving objects as sequences of regions frequently visit ed with a typical travel time.\n\nA decision tree\, named T-pattern Tree \, is built and evaluated with a formal training and test process. The t ree is learned from the Trajectory Patterns that hold a certain area and it may be used as a predictor of the next location of a new trajectory finding the best matching path in the tree. Three different best matchin g methods to classify a new moving object are proposed and their impact on the quality of prediction is studied extensively.\n\nUsing Trajectory Patterns as predictive rules has the following implications: (I) the le arning depends on the movement of all available objects in a certain are a instead of on the individual history of an object\; (II) the predictio n tree intrinsically contains the spatio-temporal properties that have e merged from the data and this allows us to define matching methods that striclty depend on the properties of such movements.\n\nIn addition\, we propose a set of other measures\, that evaluate a priori the predictive power of a set of Trajectory Patterns. This measures were tuned on a re al life case study. Finally\, an exhaustive set of experiments and resul ts on the real dataset are presented.\n SUMMARY:WhereNext: a Location Predictor on Trajectory Pattern Mining LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T115000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T113500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4924 DESCRIPTION:Given a real\, and weighted person-to-person network which ch anges over time\, what can we say about the cliques that it contains? Do the incidents of communication\, or weights on the edges of a clique fo llow any pattern? Real\, and in-person social networks have many more tr iangles than chance would dictate. As it turns out\, there are many more cliques than one would expect\, in surprising patterns.\n\nIn this pape r\, we study massive real-world social networks formed by direct contact s among people through various personal communication services\, such as Phone-Call\, SMS\, IM etc. The contributions are the following: (a) we discover surprising patterns with the cliques\, (b) we report power-law s of the weights on the edges of cliques\, (c) our real networks follow these patterns such that we can trust them to spot outliers and finally\ , (d) we propose the first utility-driven graph generator for weighted t ime-evolving networks\, which match the observed patterns. Our study foc used on three large datasets\, each of which is a different type of comm unication service\, with over one million records\, and spans several mo nths of activity. SUMMARY:Large Human Communication Networks: Patterns and a Utility-Driven Generator LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T120500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T115000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4925 DESCRIPTION:We present novel semi-supervised boosting algorithms that inc rementally build linear combinations of weak classifiers through generic functional gradient descent using both labeled and unlabeled training d ata. Our approach is based on extending information regularization frame work to boosting\, bearing loss functions that combine log loss on label ed data with the information-theoretic measures to encode unlabeled data . Even though the information-theoretic regularization terms make the op timization non-convex\, we propose simple sequential gradient descent op timization algorithms\, and obtain impressively improved results on synt hetic\, benchmark and real world tasks over supervised boosting algorith ms which use the labeled data alone and a state-of-the-art semi-supervis ed boosting algorithm. SUMMARY:Information Theoretic Regularization for Semi-Supervised Boosting LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T153500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T152000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4926 DESCRIPTION:Given multiple time sequences with missing values\, we propos e DynaMMo which summarizes\, compresses\, and finds latent variables. Th e idea is to discover hidden variables and learn their dynamics\, making our algorithm able to function even when there are missing values. We p erformed experiments on both real and synthetic datasets spanning several megabytes\, including motion capture sequences and chlorine leve ls in drinking water. \n\nWe show that our proposed DynaMMo method \n( a) can successFully learn the latent variables and their evolution\; \n( b) can provide high compression for little loss of reconstruction accura cy\; \n(c) can extract compact but powerful features for segmentation\, interpretation\, and forecasting\; \n(d) has complexity linear on the du ration of sequences. SUMMARY:DynaMMo: Mining and Summarization of Coevolving Sequences with Mi ssing Values LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T120500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T115000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4927 DESCRIPTION:Recent innovations have resulted in a plethora of social appl ications on the Web\, such as blogs\, social networks\, and community ph oto and video sharing applications. Such applications can typically be r epresented as evolving interaction graphs with nodes denoting entities a nd edges representing their interactions. The study of entities and comm unities and how they evolve in such large dynamic graphs is both importa nt and challenging.\n\nWhile much of the past work in this area has focu sed on static analysis\, more recently researchers have investigated dyn amic analysis. In this paper\, in a departure from recent efforts\, we c onsider the problem of analyzing patterns and critical events that affec t the dynamic graph from the viewpoint of a single node\, or a selected subset of nodes. Defining and extracting a relevant viewpoint neighborho od efficiently\, while also quantifying the key relationships among node s involved are the key challenges we address. We also examine the evolut ion of viewpoint neighborhoods for different entities over time to ident ify key structural and behavioral transformations that occur. SUMMARY:A Viewpoint-based Approach for Interaction Graph Analysis LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T142500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4928 DESCRIPTION:Corruption of data by class-label noise is an important pract ical concern impacting many classification problems. Studies of data cle aning techniques often assume a uniform label noise model\, however\, wh ich is seldom realized in practice. Relatively little is understood\, as to how the natural label noise distribution can be measured or simulate d. Using email spam-filtering data\, we demonstrate that class noise can have substantial content specific bias. We also demonstrate that noise detection techniques based on classifier confidence tend to identify ins tances that human assessors are likely to label in error. We show that g enre modeling can be very informative in identifying potential areas of mislabeling. Moreover\, we are able to show that genre decomposition can also be used to substantially improve spam filtering accuracy\, with ou r results outperforming the best published figures for the trec05-p1 and ceas-2008 benchmark collections. SUMMARY:Genre-based Decomposition of Email Class Noise LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T120500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T115000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4929 DESCRIPTION:We describe a recommender system in the domain of grocery sho pping. While recommender systems have been widely studied\, this is most ly in relation to leisure products (e.g. movies\, books and music) with non-repeated purchases. In grocery shopping\, however\, consumers will m ake multiple purchases of the same or very similar products more frequen tly than buying entirely new items. The proposed recommendation scheme o ffers several advantages in addressing the grocery shopping problem\, na mely: 1) a product similarity measure that suits a domain where no ratin g information is available\; 2) a basket sensitive random walk model to approximate product similarities by exploiting incomplete neighborhood i nformation\; 3) online adaptation of the recommendation based on the cur rent basket and 4) a new performance measure focusing on products that c ustomers have not purchased before or purchase infrequently. Empirical r esults benchmarking on three real-world data sets demonstrate a performa nce improvement of the proposed method over other existing collaborative filtering models. SUMMARY:Grocery Shopping Recommendations Based on Basket-Sensitive Random Walk LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T120000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T114500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4930 DESCRIPTION:In recent years\, the number of patents filed by the business enterprises in the technology industry are growing rapidly\, thus provi ding unprecedented opportunities for knowledge discovery in patent data. One important task in this regard is to employ data mining techniques t o rank patents in terms of their potential to earn money through licensi ng. Availability of such ranking can substantially reduce enterprise IP (Intellectual Property) management costs. Unfortunately\, the existing software systems in the IP domain do not address this task directly. Th rough our research\, we build a patent ranking software\, named COA (Cla im Originality Analysis) that rates a patent based on its value by measu ring the recency and the impact of the important phrases that appear in the "claims" section of a patent. Experiments show that COA produces mea ningful ranking when comparing it with other indirect patent evaluation metrics--- citation count\, patent status\, and attorney's rating. In re al-life settings\, this tool was used by beta-testers in the IBM IP depa rtment. Lawyers found it very useful in patent rating\, specifically\, i n highlighting potentially valuable patents in a patent cluster. In thi s article\, we describe the ranking techniques and system architecture o f COA. We also present the results that validate its effectiveness. SUMMARY:COA: Finding Novel Patents through Text Analysis LOCATION:La Seine C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T145000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T142500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4931 DESCRIPTION:In traditional text clustering methods\, documents are repres ented as bags of words without considering the semantic information of each document. For instance\, if two documents use different collection s of core words to represent the same topic\, they may be falsely assi gned to different clusters due to the lack of shared core words\, althou gh the core words they use are probably synonyms or semantically associa ted in other forms. The most common way to solve this problem is to enri ch document representation with the background knowledge in an ontology. There are two major issues for this approach: (1) the coverage of the o ntology is limited\, even for WordNet or Mesh\, (2) using ontology terms as replacement or additional features may cause information loss\, or i ntroduce noise. In this paper\, we present a novel text clustering metho d to address these two issues by enriching document representation with Wikipedia concept and category information. We develop two approaches\, exact match and relatedness-match\, to map text documents to Wikipedia c oncepts\, and further to Wikipedia categories. Then the text documents a re clustered based on a similarity metric which combines document conten t information\, concept information as well as category information. The experimental results using the proposed clustering framework on three d atasets (20-newsgroup\, TDT2\, and LA Times) show that clustering perfor mance improves significantly by enriching document representation with W ikipedia concepts and categories. SUMMARY:Exploiting Wikipedia as External Knowledge for Document Clusterin g LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T113500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4932 DESCRIPTION:Algorithms based on simulating stochastic flows are a simple and natural solution for the problem of clustering graphs\, but their w idespread use has been hampered by their lack of scalability and fragmen tation of output. In this article we present a multi-level algorithm for graph clustering using flows that delivers significant improvements in both quality and speed. The graph is first successively coarsened to a m anageable size\, and a small number of iterations of flow simulation is performed on the coarse graph. The graph is then successively refined\, with flows from the previous graph used as initializations for brief flo w simulations on each of the intermediate graphs. When we reach the fina l refined graph\, the algorithm is run to convergence and the high-flow regions are clustered together\, with regions without any flow forming the natural boundaries of the clusters. Extensive experimental results o n several real and synthetic datasets demonstrate the effectiveness of o ur approach when compared to state-of-the-art algorithms. SUMMARY:Scalable Graph Clustering Using Stochastic Flows: Applications to Community Discovery LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4933 DESCRIPTION:Software is a ubiquitous component of our daily life. We ofte n depend on the correct working of software systems. Due to the difficu lty and complexity of software systems\, bugs and anomalies are prevalen t. Bugs have caused billions of dollars loss\, in addition to privacy an d security threats. In this work\, we address software reliability issue s by proposing a novel method to classify software behaviors based on pa st history or runs. With the technique\, it is possible to generalize pa st known errors and mistakes to capture failures and anomalies. Our tech nique first mines a set of discriminative features capturing repetitive series of events from program execution traces. It then performs feature selection to select the best features for classification. These feature s are then used to train a classifier to detect failures. Experiments an d case studies on traces of several benchmark software systems and a rea l-life concurrency bug from MySQL server show the utility of the techniq ue in capturing failures and anomalies. On average\, our pattern-based c lassification technique outperforms the baseline approach by 24.68% in a ccuracy. SUMMARY:Classification of Software Behaviors for Failure Detection: A Dis criminative Pattern Mining Approach LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T113500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T111000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4934 DESCRIPTION:Address standardization is a very challenging task in data cl eansing. To provide better customer relationship management and busines s intelligence for customer-oriented cooperates\, millions of free-text addresses need to be converted to a standard format for data integrati on\, de-duplication and householding. Existing commercial tools usually employ lots of hand-craft\, domain-specific rules and reference data dictionary of cities\, states etc. These rules work better for the regi on they are designed. However\, rule-based methods usually require more human efforts to rewrite these rules for each new domain since address d ata are very irregular and varied with countries and regions. Supervi sed learning methods usually are more adaptable than rule-based approach es. However\, supervised methods need large-scale labeled training d ata. It is a labor-intensive and time-consuming task to build a large-s cale annotated corpus for each target domain. For minimizing human effo rts and the size of labeled training data set\, we present a free-text a ddress standardization method with latent semantic association (LaSA). LaSA model is constructed to capture latent semantic association among words from the unlabeled corpus. The original term space of the targ et domain is projected to a concept space using LaSA model at first\, th en the address standardization model is active learned from LaSA featur es and informative samples. The proposed method effectively captures th e data distribution of the domain. Experimental results on large-scale English and Chinese corpus show that the proposed method significantly enhances the performance of standardization with less efforts and train ing data. SUMMARY:Address Standardization with Latent Semantic Association LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T172000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T170500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4935 DESCRIPTION:This paper addresses the issue of unsupervised network anomal y detection. In recent years\, networks have played more and more critic al roles. Since their outages cause serious economic losses\, it is quit e significant to monitor their changes over time and to detect anomalies as early as possible. In this paper\, we specifically focus on the mana gement of the whole network. In it\, it is important to detect anomalies which make great impact on the whole network\, and the other local anom alies should be ignored. Further\, when we detect the former anomalies\, it is required to localize nodes responsible for them. It is challengin g to simultaneously perform the above two tasks \ntaking into account th e nonstationarity and strong correlations between nodes.\n\nWe propose a network anomaly detection method which resolves the above two tasks in a unified way. The key ideas of the method are: (1)construction of quant ities representing feature of a whole network and each node from the sam e input based on eigen equation compression\, and (2)incremental anomalo usness scoring based on learning the probability distribution of the qua ntities.\n\nWe demonstrate through the experimental results using two be nchmark data sets and a simulation data set that anomalies of a whole ne twork and nodes responsible for them can be detected by the proposed met hod. SUMMARY:Network Anomaly Detection based on Eigen Equation Compression LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4936 DESCRIPTION:Conditional random fields(CRFs) are a class of undirected gra phical models which have been widely used for classifying and labeling s equence data. The training of CRFs is typically formulated as an unconst rained optimization problem that maximizes the conditional likelihood. H owever\, maximum likelihood training is prone to overfitting. To address this issue\, we propose a novel constrained nonlinear optimization form ulation in which the prediction accuracy of cross-validation sets are in cluded as constraints. Instead of requiring multiple passes of training\ , the constrained formulation\nallows the cross-validation be handled in one pass of constrained optimization.\n\nThe new formulation is discont inuous\, and classical Lagrangian based constraint handling methods are not applicable. A new constrained optimization algorithm based on the re cently proposed extended saddle point theory is developed to learn the c onstrained CRF model. Experimental results on gene and stock-price predi ction tasks show\nthat the constrained formulation is able to significan tly improve the generalization ability of CRF training. SUMMARY:Constrained Optimization for Validation-Guided Conditional Random Field Learning LOCATION:Foyer Rives de Seine - Pont des Arts A & B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T151500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T145000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4937 DESCRIPTION:Search logs\, which contain rich and up-to-date information a bout users' needs and preferences\, have become a critical data source f or search engines. Recently\, more and more data-driven applications are being developed in search engines based on search logs\, such as query suggestion\, keyword bidding\, and dissatisfactory query analysis. In th is paper\, by observing that many data-driven applications in search eng ines highly rely on online mining of search logs\, we develop an OLAP sy stem on search logs which serves as an infrastructure supporting various data-driven applications. An empirical study using real data of over tw o billion query sessions demonstrates the usefulness and feasibility of our design. SUMMARY:OLAP on Search Logs: An Infrastructure Supporting Data-Driven App lications in Search Engines LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T144000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T142500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4938 DESCRIPTION:Lately there exist increasing demands for online abnormality monitoring over trajectory streams\, which are obtained from moving obje ct tracking devices. This problem is challenging due to the requirement of high speed data processing within limited space cost. In this paper\, we present a novel framework for monitoring anomalies over continuous t rajectory streams. First\, we illustrate the importance of distance-base d anomaly monitoring over moving object trajectories. Then\, we utilize the local continuity characteristics of trajectories to build local clus ters upon trajectory streams and monitor anomalies via efficient pruning strategies. Finally\, we propose a piecewise metric index structure to reschedule the joining order of local clusters to further reduce the tim e cost. Our extensive experiments demonstrate the effectiveness and effi ciency of our methods. SUMMARY:Efficient Anomaly Monitoring Over Moving Object Trajectory Stream s LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T142500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4939 DESCRIPTION:Linear classifiers have been shown to be effective for many d iscrimination tasks. Irrespective of the learning algorithm itself\, th e final classifier has a weight to multiply by each feature. This sugges ts that ideally each input feature should be linearly correlated with th e target variable (or anti-correlated)\, whereas raw features may be hig hly non-linear. In this paper\, we attempt to re-shape each input featu re so that it is appropriate to use with a linear weight and to scale th e different features in proportion to their predictive value. We demons trate that this pre-processing is beneficial for linear SVM classifiers on a large benchmark of text classification tasks as well as UCI dataset s. SUMMARY:Feature Shaping for Linear SVM Classifiers LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T170500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T165000 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4940 DESCRIPTION:As the number and size of large timestamped collections (e.g. sequences of digitized newspapers\, periodicals\, blogs) increase\, the problem of efficiently indexing and searching such data becomes more im portant. Term burstiness has been extensively researched as a mechanism to address event detection in the context of such collections. In this p aper\, we explore how burstiness information can be further utilized to enhance the search process. We present a novel approach to model the bur stiness of a term\, using discrepancy theory concepts. This allows us to build a parameter-free\, linear-time approach to identify the time inte rvals of maximum burstiness for a given term.\nFinally\, we describe the first burstiness-driven search framework and thoroughly evaluate our ap proach in the context of different scenarios. SUMMARY:On Burstiness-aware Search for Document Sequences LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T122000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T120500 DTSTAMP;VALUE=DATE-TIME:20120516T153738Z UID:4941 DESCRIPTION:Recently many data types arising from data mining and Web sea rch applications can be modeled as bipartite graphs. Examples include qu eries and URLs in query logs\, and authors and papers in scientific lite rature. However\, one of the issues is that previous algorithms only con sider the content and link information from one side of the bipartite gr aph. There is a lack of constraints to make sure the final relevance of the score propagation on the graph\, as there are many noisy edges withi n the bipartite graph. In this paper\, we propose a novel and general Co -HITS algorithm to incorporate the bipartite graph with the content info rmation from both sides as well as the constraints of relevance. Moreove r\, we investigate the algorithm based on two frameworks\, including the iterative and the regularization frameworks\, and illustrate the genera lized Co-HITS algorithm from different views. For the iterative framewor k\, it contains HITS and personalized PageRank as special cases. In the regularization framework\, we successfully build a connection with HITS\ , and develop a new cost function to consider the direct relationship be tween two entity sets\, which leads to a significant improvement over th e baseline method. To illustrate our methodology\, we apply the Co-HITS algorithm\, with many different settings\, to the application of query s uggestion by mining the AOL query log data. Experimental results demonst rate that CoRegu-0.5 (i.e.\, a model of the regularization framework) ac hieves the best performance with consistent and promising improvements. SUMMARY:A Generalized Co-HITS Algorithm and Its Application to Bipartite Graphs LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T121500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T115000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4942 DESCRIPTION:Advanced analysis of data streams is quickly becoming a key a rea of data mining research as the number of applications demanding such processing increases. Online mining when such data streams evolve over time\, that is when concepts drift or change completely\, is becoming on e of the core issues. When tackling non-stationary concepts\, ensembles of classifiers have several advantages over single classifier methods: t hey are easy to scale and parallelize\, they can adapt to change quickly by pruning under-performing parts of the ensemble\, and they therefore usually also generate more accurate concept descriptions.\n\nThis paper proposes a new experimental data stream framework for studying concept d rift\, and two new variants of Bagging: ADWIN Bagging and Adaptive-Size Hoeffding Tree (ASHT) Bagging. Using the new experimental framework\, a n evaluation study on synthetic and real-world datasets comprising up to ten million examples shows that the new ensemble methods perform very w ell compared to several known methods. SUMMARY:New ensemble methods for evolving data streams LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T145500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T144000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4943 DESCRIPTION:How can we automatically spot all outstanding observations in a data set? This question arises in a large variety of applications\, e .g. in economy\, biology and medicine. Existing approaches to outlier de tection suffer from one or more of the following drawbacks: The results of many methods strongly depend on suitable parameter settings being ve ry difficult to estimate without background\nknowledge on the data\, e.g . the minimum cluster size or the number of desired outliers. Many metho ds implicitly assume Gaussian or uniformly distributed data\, and/or the ir result is difficult to interpret. To cope with these problems\, we pr opose CoCo\, a technique for parameter-free outlier detection. The basic idea of our technique relates outlier detection to data compression: Ou tliers are objects which can not be effectively compressed given the dat a set. To avoid the assumption of a certain data distribution\, CoCo rel ies on a very general data model combining\nthe Exponential Power Distri bution with Independent Components. We define an intuitive outlier facto r based on the principle of the Minimum Description Length together with an novel algorithm for outlier detection. An extensive experimental eva luation on synthetic and real world data demonstrates the benefits of ou r technique. Availability: The source code of CoCo and the data sets use d in the experiments are available at:\nhttp://www.dbs.ifi.lmu.de/Forsch ung/KDD/Boehm/CoCo. SUMMARY:CoCo: Coding Cost for Parameter-free Outlier Detection LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T112500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T111000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4944 DESCRIPTION:Time series prediction is an important issue in a wide range of areas. There are various real world processes whose states vary conti nuously\, and those processes may have influences on each other. If the past information of one process X improves the predictability of another process Y\, X is said to have a causal influence on Y. In order to make good predictions\, it is necessary to identify the appropriate causal r elationships. In addition\, the processes to be modeled may include symb olic data as well as numerical data. Therefore\, it is important to deal with symbolic and numerical time series seamlessly when attempting to d etect causality.\n\nIn this paper\, we propose a new method for quantify ing the strength of the causal influence from one time series to another . The proposed method can represent the strength of causality as the num ber of bits\, whether each of two time series is symbolic or numerical. The proposed method can quantify causality even from a small number of s amples. In addition\, we propose structuring and modeling methods for mu ltivariate time series using causal relationships of two time series. Ou r structuring and modeling methods can also deal with data sets which in clude both types of time series. Experimental results demonstrate that o ur methods can perform well even if the number of samples is small. SUMMARY:Causality Quantification and Its Applications: Structuring and Mo deling of Multivariate Time Series LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T173500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T172000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4945 DESCRIPTION:One common predictive modeling challenge occurs in text minin g problems is that the training data and the operational (testing) data are drawn from different underlying distributions. This poses a great di fficulty for many statistical learning methods. However\, when the distr ibution in the source domain and the target domain are not identical but related\, there may exist a shared concept space to preserve the relati on. Consequently a good feature representation can encode this concept s pace and minimize the distribution gap. To formalize this intuition\, we propose a domain adaptation method that parameterizes this concept spac e by linear transformation under which we explicitly minimize the distri bution difference between the source domain with sufficient labeled data and target domains with only unlabeled data\, while at the same time mi nimizing the empirical loss on the labeled data in the source domain. An other characteristic of our method is its capability for considering mul tiple classes and their interactions simultaneously. We have conducted e xtensive experiments on two common text mining problems\, namely\, infor mation extraction and document classification to demonstrate the effecti veness of our proposed method. SUMMARY:Extracting Discriminative Concepts for Domain Adaptation in Text Mining LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T165500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T164000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4946 DESCRIPTION:Research in relational data mining has two major directions: finding global models of a relational database and the discovery of loca l relational patterns within a database. While relational patterns show how attribute values co-occur in detail\, their huge numbers hamper thei r usage in data analysis. Global models\, on the other hand\, only provi de a summary of how different tables and their attributes relate to each other\, lacking detail of what is going on at the local level.\n\nIn th is paper we introduce a new approach that combines the positive properti es of both directions: it provides a detailed description of the complet e database using a small set of patterns. More in particular\, we utilis e a rich pattern language and show how a database can be encoded by such patterns. Then\, based on the MDLprinciple\, the novel RDB-KRIMP algori thm selects the set of patterns that allows for the most succinct encodi ng of the database. This set\, the code table\, is a compact description of the database in terms of local relational patterns. We show that thi s resulting set is very small\, both in terms of database size and in nu mber of its local relational patterns: a reduction of up to 4 orders of magnitude is attained. SUMMARY:Characteristic Relational Patterns LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4947 DESCRIPTION:Website traffic varies through time in consistent and predict able ways\, with highest traffic in the middle of the day. When providi ng media content to visitors\, it is important to present repeat visitor s with new content so that they keep coming back. In this paper we pres ent an algorithm to balance the need to keep a website fresh with new co ntent with the desire to present the best content to the most visitors a t times of peak traffic. We formulate this as the media scheduling prob lem\, where we attempt to maximize total clicks\, given the overall traf fic pattern and the time varying clickthrough rates of available media c ontent. We present an efficient algorithm to perform this scheduling un der certain conditions and apply this algorithm to real data obtained fr om server logs\, showing evidence of significant improvements in traffic from our algorithmic schedules. Finally\, we analyze the click data\, presenting models for why and how the clickthrough rate for new content declines as it ages. SUMMARY:Optimizing Web Traffic via the Media Scheduling Problem LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4948 DESCRIPTION:In this study\, we formalize a multi-focal learning problem\, where training data are partitioned into several different focal groups and the prediction model will be learned within each focal group. The m ulti-focal learning problem is motivated by numerous real-world learning applications. For instance\, for the same type of problems encountered in a customer service center\, the problem descriptions from different c ustomers can be quite different. The experienced customers usually give more precise and focused descriptions about the problem. In contrast\, t he inexperienced customers usually provide more diverse descriptions. In this case\, the examples from the same class in the training data can b e naturally in different focal groups. As a result\, it is necessary to identify those natural focal groups and exploit them for learning at dif ferent focuses. The key developmental challenge is how to identify those focal groups in the training data. As a case study\, we exploit multi-f ocal learning for profiling problems in customer service centers. The re sults show that multifocal learning can significantly boost the learning accuracies of existing learning algorithms\, such as Support Vector Mac hines (SVMs)\, for classifying customer problems. SUMMARY:Multi-focal Learning and Its Application to Customer Service Supp ort LOCATION:La Seine C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4949 DESCRIPTION:Network data is ubiquitous\, encoding collections of relation ships between entities such as people\, places\, genes\, or corporatio ns. While many resources for networks of interesting entities are em erging\, most of these can only annotate connections in a limited fash ion. Although relationships between entities are rich\, it is impracti cal to manually devise complete characterizations of these relationshi ps for every pair of entities on large\, real-world corpora.\n\nIn thi s paper we present a novel probabilistic topic model to analyze text c orpora and infer descriptions of its entities and of relationships bet ween those entities. We develop variational methods for performing ap proximate inference on our model and demonstrate that our model can be p ractically deployed on large corpora such as Wikipedia. We show quali tatively and quantitatively that our model can construct and annotate graphs of relationships and make useful predictions. SUMMARY:Connections between the Lines: Augmenting Social Networks with Te xt LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4950 DESCRIPTION:Probabilistic frequent itemset mining in uncertain transactio n databases semantically and computationally differs from traditional te chniques applied to standard certain transaction databases. The consid eration of existential uncertainty of item(sets)\, indicating the probab ility that an item(set) occurs in a transaction\, makes traditional tech niques inapplicable. In this paper\, we introduce new probabilistic form ulations of frequent itemsets based on possible world semantics. In this probabilistic context\, an itemset X is called frequent if the probabil ity that X occurs in at least minSup transactions is above a given thres hold ?. To the best of our knowledge\, this is the first approach addres sing this problem under possible worlds semantics. In consideration of t he probabilistic formulations\, we present a framework which is able to solve the Probabilistic Frequent Itemset Mining (PFIM) problem efficient ly. An extensive experimental evaluation investigates the impact of our proposed techniques and shows that our approach is orders of magnitude f aster than straight-forward approaches. SUMMARY:Probabilistic Frequent Itemset Mining in Uncertain Databases LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T145000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T142500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4951 DESCRIPTION:Max-margin Markov networks (M3N) have shown great promise in structured prediction and relational learning. Due to the KKT conditions \, the M3N enjoys dual sparsity. However\, the existing M3N formulation does not enjoy primal sparsity\, which is a desirable property for selec ting significant features and reducing the risk of over-fitting. In this paper\, we present an l1-norm regularized max-margin Markov network (l1 -M3N)\, which enjoys dual and primal sparsity simultaneously. To learn a n l1-M3N\, we present three methods including projected sub-gradient\, c utting-plane\, and a novel EM-style algorithm\, which is based on an equ ivalence between l1-M3N and an adaptive M3N. We perform extensive empiri cal studies on both synthetic and real data sets. Our experimental resul ts show that: (1) l1-M3N can effectively select significant features\; ( 2) l1-M3N can perform as well as the pseudo-primal sparse Laplace M3N in prediction accuracy\, while consistently outperforms other competing me thods that enjoy either primal or dual sparsity\; and (3) the EM-algorit hm is more robust than the other two in prediction accuracy and time eff iciency. SUMMARY:Primal Sparse Max-Margin Markov Networks LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T120000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T114500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4952 DESCRIPTION:Tag recommendation is the task of predicting a personalized l ist of tags for a user given an item. This is important for many website s with tagging capabilities like last.fm or delicious. In this paper\, w e propose a method for tag recommendation based on tensor factorization (TF). In contrast to other TF methods like higher order singular value d ecomposition (HOSVD)\, our method RTF (`ranking with tensor factorizatio n') directly optimizes the factorization model for the best personalized ranking. RTF handles missing values and learns from pairwise ranking co nstraints. Our optimization criterion for TF is motivated by a detailed analysis of the problem and of interpretation schemes for the observed d ata in tagging systems. In all\, RTF directly optimizes for the actual p roblem using a correct interpretation of the data. We provide a gradient descent algorithm to solve our optimization problem. We also provide an improved learning and prediction method with runtime complexity analysi s for RTF. The prediction runtime of RTF is independent of the number of observations and only depends on the factorization dimensions. Besides the theoretical analysis\, we empirically show that our method outperfor ms other state-of-the-art tag recommendation methods like FolkRank\, Pag eRank and HOSVD both in quality and prediction runtime. SUMMARY:Learning Optimal Ranking with Tensor Factorization for Tag Recomm endation LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T145500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T144000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4953 DESCRIPTION:In large social networks\, nodes (users\, entities) are influ enced by others for various reasons. For example\, the colleagues have s trong influence on one's work\, while the friends have strong influence on one's daily life. How to differentiate the social influences from dif ferent angles(topics)? How to quantify the strength of those social infl uences? How to estimate the model on real large networks?\n\nTo address these fundamental questions\, we propose Topical Affinity Propagation (T AP) to model the topic-level social influence on large networks. In part icular\, TAP can take results of any topic modeling and the existing net work structure to perform topic-level influence propagation. With the he lp of the influence analysis\, we present several important applications on real data sets such as 1) what are the representative nodes on a giv en topic? 2) how to identify the social influences of neighboring nodes on a particular node? \n\nTo scale to real large networks\, TAP is desig ned with efficient distributed learning algorithms that is implemented a nd tested under the Map-Reduce framework. We further present the common characteristics of distributed learning algorithms for Map-Reduce. Final ly\, we demonstrate the effectiveness and efficiency of TAP on real larg e data sets. SUMMARY:Social Influence Analysis in Large-scale Networks LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T113500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4954 DESCRIPTION:Opinion mining became an important topic of study in recent y ears due to its wide range of applications. There are also many companie s offering opinion mining services. One problem that has not been studie d so far is the assignment of entities that have been talked about in ea ch sentence. Let us use forum discussions about products as an example t o make the problem concrete. In a typical discussion post\, the author m ay give opinions on multiple products and also compare them. The issue i s how to detect what products have been talked about in each sentence. I f the sentence contains the product names\, they need to be identified. We call this problem entity discovery. If the product names are not expl icitly mentioned in the sentence but are implied due to the use of prono uns and language conventions\, we need to infer the products. We call th is problem entity assignment. These problems are important because witho ut knowing what products each sentence talks about the opinion mined fro m the sentence is of little use. In this paper\, we study these problems and propose two effective methods to solve the problems. Entity discove ry is based on pattern discovery and entity assignment is based on minin g of comparative sentences. Experimental results using a large number of forum posts demonstrate the effectiveness of the technique. Our system has also been successfully tested in a commercial setting. SUMMARY:Entity Discovery and Assignment for Opinion Mining Applications LOCATION:Louis Armstrong A\,B\,C+D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T121500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4955 DESCRIPTION:Correlated or discriminative pattern mining is concerned with finding the highest scoring patterns w.r.t. a correlation measure (such as information gain). By reinterpreting correlation measures in ROC spa ce and formulating correlated itemset mining as a constraint programming problem\, we obtain new theoretical insights with practical benefits. M ore specifically\, we contribute 1) an improved bound for correlated ite mset miners\, 2) a novel iterative pruning algorithm to exploit the boun d\, and 3) an adaptation of this algorithm to mine all itemsets on the c onvex hull in ROC space. The algorithm does not depend on a minimal freq uency threshold and is shown to outperform several alternative approache s by orders of magnitude\, both in runtime and in memory requirements. SUMMARY:Correlated Itemset Mining in ROC Space: A Constraint Programming Approach LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T151000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T145500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4956 DESCRIPTION:Given a task T\, a pool of individuals X with different skill s\, and a social network G that captures the compatibility among these i ndividuals\, we study the problem of finding X\, a subset of X\, to perf orm the task. We call this the Team Formation problem. We require that m embers of X not only meet the skill requirements of the task\, but can a lso work effectively together as a team. We measure effectiveness using the communication cost incurred by the subgraph in G that only involves X. We study two variants of the problem for two different communication- cost functions\, and show that both variants are NP-hard. We explore the ir connections with existing combinatorial problems and give novel algor ithms for their solution. To the best of our knowledge\, this is the fir st work to consider the Team Formation problem in the presence of a soci al network of individuals. Experiments on the DBLP dataset show that our framework works well in practice and gives useful and intuitive results . SUMMARY:Finding a Team of Experts in Social Networks LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4957 DESCRIPTION:Currently\, the most significant line of defense against malw are is anti-virus products which focus on authenticating valid software from a white list\, blocking invalid software from a black list\, and ru nning any unknown software (i.e.\, the gray list) in a controlled manner . The gray list\, containing unknown software programs which could be ei ther normal or malicious\, is usually authenticated or rejected manually by virus analysts. Unfortunately\, along with the development of the ma lware writing techniques\, the number of file samples in the gray list t hat need to be analyzed by virus analysts on a daily basis is constantly increasing. In this paper\, we develop an intelligent file scoring syst em (IFSS for short) for malware detection from the gray list by an ensem ble of heterogeneous base-level classifiers derived by different learnin g methods\, using different feature representations on dynamic training sets. To the best of our knowledge\, this is the first work of applying such ensemble methods for malware detection. IFSS makes it practical for \nvirus analysts to identify malware samples from the huge gray list and improves the detection ability of anti-virus software. It has already b een incorporated into the scanning tool of Kingsoft's Anti-Virus softwar e. The case studies on large and real daily collection of the gray list illustrate that the detection ability and efficiency of our IFSS system outperforms other popular scanning tools such as NOD32 and Kaspersky. SUMMARY:Intelligent File Scoring System for Malware Detection from the Gr ay List LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4958 DESCRIPTION:For implementing content management solutions and enabling ne w applications associated with data retention\, regulatory compliance\, and litigation issues\, enterprises need to develop advanced analytics t o uncover relationships among the documents\, e.g.\, content similarity\ , provenance\, and clustering. In this paper\, we evaluate the performa nce of four syntactic similarity algorithms. Three algorithms are based on Broder's ``shingling'' technique while the fourth algorithm employs a more recent approach\, ``content-based chunking''. For our experiments\ , we use a specially designed corpus of documents that includes a set of ``similar'' documents with a controlled number of modifications. Our pe rformance study reveals that the similarity metric of all four algorithm s is highly sensitive to settings of the algorithms' parameters: slidin g window size and fingerprint sampling frequency. We identify a useful r ange of these parameters for achieving good practical results\, and comp are the performance of the four algorithms in a controlled environment. We validate our results by applying these algorithms to finding near-du plicates in two large collections of HP technical support documents. SUMMARY:Applying Syntactic Similarity Algorithms for Enterprise Informati on Management LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T121500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T120000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4959 DESCRIPTION:Fault-tolerant frequent itemsets (FTFI) are variants of frequ ent itemsets for representing and discovering generalized knowledge. How ever\, despite growing interest in this field\, no previous approach min es proportional FTFIs with their exact support (FT-support).\n\nThis pro blem is difficult because of two concerns: (a) non anti-monotonic proper ty of FT-support when relaxation is proportional\, and (b) difficulty in computing FT-support. Previous efforts on this problem either simplify the general problem by adding constraints\, or provide approximate solut ions without any error guarantees.\n\nIn this paper\, we address these c oncerns in the general FTFI mining problem. We limit the search space by providing provably correct anti monotone bounds for FT-support and deve lop practically efficient means of achieving them. Besides\, we also pro vide an efficient and exact FT-support counting procedure.\n\nExtensive experiments using real datasets validate that our solution is reasonably efficient for completely mining FTFIs. Implementations for the algorith ms are available from http://www3.ntu.edu.sg/home/asvivek/pubs/ftfim09/. SUMMARY:Towards Efficient Mining of Proportional Fault-Tolerant Frequent Itemsets LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4960 DESCRIPTION:This paper explores an important and relatively unstudied qua lity measure of a sponsored search advertisement: bounce rate. The bounc e rate of an ad can be informally defined as the fraction of users who c lick on the ad but almost immediately move on to other tasks. A high bou nce rate can lead to poor advertiser return on investment\, and suggests search engine users may be having a poor experience follow ing the clic k. In this paper\, we first provide quantitative analysis showing that b ounce rate is an effective measure of user satisfaction. We then address the question\, can we predict bounce rate by analyzing the features of the advertisement? An affirmative answer would allow advertisers and sea rch engines to predict the effectiveness and quality of advertisements b efore they are shown. We propose solutions to this problem involving lar ge-scale learning methods that leverage features drawn from ad creatives in addition to their keywords and landing pages. SUMMARY:Predicting Bounce Rates in Sponsored Search Advertisements LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T111000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4961 DESCRIPTION:The importance of event logs\, as a source of information in systems and network management cannot be overemphasized. With the ever i ncreasing size and complexity of today's event logs\, the task of analyz ing event logs has become cumbersome to carry out manually. For this rea son recent research has focused on the automatic analysis of these log f iles. In this paper we present IPLoM (Iterative Partitioning Log Mining) \, a novel algorithm for the mining of clusters from event logs. Through a 3-Step hierarchical partitioning process IPLoM partitions log data in to its respective clusters. In its 4th and final stage IPLoM produces cl uster descriptions or line formats for each of the clusters produced. Un like other similar algorithms IPLoM is not based on the Apriori algorith m and it is able to find clusters in data whether or not its instances a ppear frequently. Evaluations show that IPLoM outperforms the other algo rithms statistically significantly\, and it is also able to achieve an a verage F-Measure performance 78% when the closest other algorithm achiev es an F-Measure performance of 10%. SUMMARY:Clustering Event Logs Using Iterative Partitioning LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T115000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T113500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4962 DESCRIPTION:Modern communication networks generate massive volume of oper ational event data\, e.g.\, alarm\, alert\, and metrics\, which can be u sed by a network management system (NMS) to diagnose potential faults. I n this work\, we introduce a new class of indexable {\\it fault signatu res} that encode temporal evolution of events generated by a network fa ult as well as topological relationships among the nodes where these eve nts occur. We present an efficient learning algorithm to extract such fa ult signatures from noisy historical event data\, and with the help of novel space-time indexing structures\, we show how to perform efficient\ , online signature matching. We provide results from extensive experimen tal studies to explore the efficacy of our approach and point out potent ial applications of such signatures for many different types of networks including social and information networks. SUMMARY:Learning\, Indexing\, and Diagnosing Network Faults LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T152000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T150500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4963 DESCRIPTION:All Netflix Prize algorithms proposed so far are prohibitivel y costly for large-scale production systems. In this paper\, we describe an efficient dataflow implementation of a collaborative filtering (CF) solution to the Netflix Prize problem [1] based on weighted coclustering [5]. The dataflow library we use facilitates the development of sophist icated parallel programs designed to fully utilize commodity multicore h ardware\, while hiding traditional difficulties such as queuing\, thread ing\, memory management\, and deadlocks. The dataflow CF implementation first compresses the large\, sparse training dataset into co-clusters. T hen it generates recommendations by combining the average ratings of the co-clusters with the biases of the users and movies. When configured to identify 20x20 co-clusters in the Netflix training dataset\, the implem entation predicted over 100 million ratings in 16.31 minutes and achieve d an RMSE of 0.88846 without any fine-tuning or domain knowledge. This i s an effective real-time prediction runtime of 9.7 ºs per rating which is far superior to previously reported results. Moreover\, the implement ed co-clustering framework supports a wide variety of other large-scale data mining applications and forms the basis for predictive modeling on large\, dyadic datasets [4\, 7]. SUMMARY:Pervasive Parallelism in Data Mining: Dataflow solution to Co-clu stering Large and Sparse Netflix Data LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T173500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T172000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4964 DESCRIPTION:In this paper\, we consider the problem of combining link and content analysis for community detection from networked data\, such as paper citation networks and Word Wide Web. Most existing approaches comb ine link and content information by a generative model that generates bo th links and contents via a shared set of community memberships. These g enerative models have some shortcomings in that they failed to consider additional factors that could affect the\ncommunity memberships and isol ate the contents that are irrelevant to community memberships. To explic itly address these shortcomings\, we propose a discriminative model for combining the link and content analysis for community detection. First\, we propose a conditional model for link analysis and in the model\, we introduce hidden variables to explicitly model the popularity of nodes. Second\, to alleviate the impact of irrelevant content attributes\, we d evelop a discriminative model for content analysis. These two models are unified seamlessly via the community memberships. We present efficient algorithms to solve the related optimization problems based on bound opt imization and alternating projection. Extensive experiments with benchma rk data sets show that the proposed framework significantly outperforms the state-of-the-art approaches for combining link and content analysis for community detection. SUMMARY:Combining Link and Content for Community Detection: A Discriminat ive Approach LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4965 DESCRIPTION:There is a growing number of service providers that a consume r can interact with over the web to learn their service terms. The servi ce terms\, such as price and time to completion of the service\, depend on the consumer's particular specifications. For instance\, a printing s ervices provider would need from its customers specifications such as th e size of paper\, type of ink\, proofing and perforation. In a few secto rs\, there exist marketplace sites that provide consumers with specifica tions forms\, which the consumer can fill out to learn the service terms of multiple service providers. Unfortunately\, there are only a few suc h marketplace sites\, and they cover a few sectors.\n\nAt HP Labs\, we a re working towards building a universal marketplace site\, i.e.\, a mark etplace site that covers thousands of sectors and hundreds of providers per sector. One issue in this domain is the automated discovery/retrieva l of the specifications for each sector. We address it through extractin g and analyzing content from the websites of the service providers liste d in business directories. The challenge is that each service provider i s often listed under multiple service categories in a business directory \, making it infeasible to utilize standard supervised learning techniqu es. We address this challenge through employing a multilabel statistical clustering approach within an expectation-maximization framework. We im plement our solution to retrieve specifications for 3000 sectors\, repre senting more than 300\,000 service providers. We discuss our results wit hin the context of the services needed to design a marketing campaign fo r a small business. SUMMARY:Towards a Universal Marketplace over the Web: Statistical Multi-l abel Classification of Service Provider Forms with Simulated Annealing LOCATION:Louis Armstrong A\,B\,C+D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4966 DESCRIPTION:Anomalous windows are the contiguous groupings of data points . In this paper\, we propose an approach for discovering anomalous windo ws using Scan Statistics for Linear Intersecting Paths (SSLIP). A linear path refers to a path represented by a line with a single dimensional s patial coordinate marking an observation point. Our approach for discove ring anomalous windows along linear paths comprises of the following dis tinct steps: (a) Cross Path Discovery: where we identify a subset of int ersecting paths to be considered\, (b) Anomalous Window Discovery: where we outline three order invariant algorithms\, namely SSLIP\, Brute Forc e-SSLIP and Central Brute Force-SSLIP\, for the traversal of the cross p aths to identify varying size directional windows along the paths. For i dentifying an anomalous window we compute an unusualness metric\, in the form of a likelihood ratio to indicate the degree of unusualness of thi s window with respect to the rest of the data. We identify the window wi th the highest likelihood ratio as our anomalous window\, and (c) Monte Carlo Simulations: to ascertain whether this window is truly anomalous a nd not just a random occurrence we perform hypothesis testing by computi ng a p-value using Monte Carlo Simulations. We present extensive experim ental results in real world accident datasets for various highways with known issues(code and data available from [27]\, [21]). Our results show that our approach indeed is effective in identifying anomalous traffic accident windows along multiple intersecting highways. SUMMARY:Anomalous Window Discovery through Scan Statistics for Linear Int ersecting Paths (SSLIP) LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4967 DESCRIPTION:Classifying nodes in networks is a task with a wide range of applications. It can be particularly useful in anomaly and fraud detecti on. Many resources are invested in the task of fraud detection due to th e high cost of fraud\, and being able to automatically detect potential fraud quickly and precisely allows human investigators to work more eff iciently. Many data analytic schemes have been put into use\; however\, schemes that bolster link analysis prove promising. This work builds upo n the belief propagation algorithm for use in detecting collusion and ot her fraud schemes. We propose an algorithm called SNARE (Social Network Analysis for Risk Evaluation). By allowing one to use domain knowledge a s well as link knowledge\, the method was very successful for pinpointin g misstated accounts in our sample of general ledger data\, with a signi ficant improvement over the default heuristic in true positive rates\, a nd a lift factor of up to 6.5 (more than twice that of the default heuri stic). We also apply SNARE to the task of graph labeling in general on p ublicly-available datasets. We show that with only some information abou t the nodes themselves in a network\, we get surprisingly high accuracy of labels. Not only is SNARE applicable in a wide variety of domains\, b ut it is also robust to the choice of parameters and highly scalable li nearly with the number of edges in a graph. SUMMARY:SNARE: A Link Analytic System for Graph Labeling and Risk Detecti on LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4968 DESCRIPTION:In this paper\, we consider a novel scheme referred to as Car tesian contour to concisely represent the collection of frequent itemset s. Different from the existing works\, this scheme provides a complete v iew of these itemsets by covering the entire collection of them. More in terestingly\, it takes a first step in deriving a generative view of the frequent pattern formulation\, i.e.\, how a small number of patterns in teract with each other and produce the complexity of frequent itemsets. We perform a theoretical investigation of the concise representation pro blem and link it to the biclique set cover problem and prove its NP-hard ness. We develop a novel approach utilizing the technique developed in f requent itemset mining\, set cover\, and max k-cover to approximate the minimal biclique set cover problem. In addition\, we consider several he uristic techniques to speedup the construction of Cartesian contour. The detailed experimental study demonstrates the effectiveness and efficien cy of our approach. SUMMARY:Cartesian Contour: A Concise Representation for a Collection of F requent Sets LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4969 DESCRIPTION:Given a quarter of petabyte click log data\, how can we estim ate the relevance of each URL for a given query? In this paper\, we prop ose the Bayesian Browsing Model (BBM)\, a new modeling technique with fo llowing advantages:\n (a) it does exact inference\;\n (b) it is single-p ass and parallelizable\;\n (c) it is effective.\n\nWe present two sets o f experiments to test model effectiveness and efficiency. On the first s et of over 50 million search instances of 1.1 million distinct queries\, BBM outperforms the state-of-the-art competitor by 29.2% in log-likelih ood while being 57 times faster. On the second click-log set\, spanning a quarter of petabyte data\, we showcase the scalability of BBM: we impl emented it on a commercial MapReduce cluster\, and it took only 3 hours to compute the relevance for 1.15 billion distinct query-URL pairs. SUMMARY:BBM: Bayesian Browsing Model from Petabyte-scale Data LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T115000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T113500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4970 DESCRIPTION:A recent study by two prominent finance researchers\, Fama an d French\, introduces a new framework for studying risk vs. return: the migration of stocks across size-value portfolio space. Given the financ ial events of 2008\, this first attempt to disentangle the relationships between migration behavior and stock returns is especially timely. The ir work\, however\, derives results only for market segments\, not indiv idual companies\, and only for one-year moves. Thus\, we see a new chall enge for financial data mining: how to capture and categorize the migrat ion of individual companies\, and how such behavior affects their return s.\n\nWe propose a novel data mining approach to study the multi-year mo vement of individual companies. Specifically\, we address the question: ``How does one discover frequent migration patterns in the stock market? '' We present a new trajectory mining algorithm to discover migration mo tifs in financial markets. Novel features of this algorithm are its hand ling of approximate pattern matching through a graph theoretical method\ , maximal clique identification\, and incorporation of temporal and spat ial constraints. We have performed a detailed study of the NASDAQ\, NYSE \, and AMEX stock markets\, over a 43-year span. We successfully find mi gration motifs that confirm existing finance theories and other motifs t hat may lead to new financial models. SUMMARY:Migration Motif: A Spatial-Temporal Pattern Mining Approach for F inancial Markets LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T114500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4971 DESCRIPTION:This paper tackles the problem of summarizing frequent itemse ts. We observe that previous notions of summaries cannot be directly use d for analyzing frequent itemsets. In order to be used for analysis\, on e requirement is that the analysts should be able to browse all frequent itemsets by only having the summary.\n\nFor this purpose\, we propose t o build the summary based upon a novel formulation\, conditional profile (or c-profile). Several features of our proposed summary are: (1) each profile in the summary can be analyzed independently\, (2) it provides e rror guarantee (e-adequate)\, and (3) it produces no false positives or false negatives.\n\nHaving the formulation\, the next challenge is to pr oduce the most concise summary which satisfies the requirement. In this paper\, we also designed an algorithm which is both effective and effici ent for this task. The quality of our approach is justified by extensive experiments.\n\nThe implementations for the algorithms are available fr om www.ntu.edu.sg/home/asvivek/pubs/cprofile09. SUMMARY:CP-Summary: A Concise Representation for Browsing Frequent Itemse ts LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4972 DESCRIPTION:In our work\, we address the problem of modeling social netwo rk generation which explains both link and group formation. Recent studi es on social network evolution propose generative models which capture t he statistical properties of real-world networks related only to node-to -node link formation. We propose a novel model which captures the co-evo lution of social and affiliation networks. We provide surprising insight s into group formation based on observations in several real-world netwo rks\, showing that users often join groups for reasons other than their friends. Our experiments show that the model is able to capture both the newly observed and previously studied network properties. This work is the first to propose a generative model which captures the statistical p roperties of these complex networks. The proposed model facilitates cont rolled experiments which study the effect of actors' behavior on the evo lution of affiliation networks\, and it allows the generation of realist ic synthetic datasets. SUMMARY:Co-evolution of Social and Affiliation Networks LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4973 DESCRIPTION:Effective diagnosis of Alzheimer's disease (AD)\, the most co mmon type of dementia in elderly patients\, is of primary importance in biomedical research. Recent studies have demonstrated that AD is closely related to the structure change of the brain network\, i.e.\, the conne ctivity among different brain regions. The connectivity patterns will pr ovide useful imaging-based biomarkers to distinguish\nNormal Controls (N C)\, patients with Mild Cognitive Impairment (MCI)\, and patients with A D. In this paper\, we investigate the sparse inverse covariance estimati on technique for identifying the connectivity among different brain regi ons. In particular\, a novel algorithm based on the block coordinate des cent approach is proposed for the direct estimation of the inverse covar iance matrix. One appealing feature of the proposed algorithm is that it allows the user feedback (e.g.\, prior domain knowledge) to be incorpor ated into the estimation process\, while the connectivity patterns can b e discovered automatically. We apply the proposed algorithm to a collect ion of FDG-PET images from 232 NC\, MCI\, and AD subjects. Our experimen tal results demonstrate that the proposed algorithm is promising in reve aling the brain region connectivity differences among these groups. SUMMARY:Mining Brain Region Connectivity for Alzheimer's Disease Study vi a Sparse Inverse Covariance Estimation LOCATION:La Seine C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T171000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T165500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4974 DESCRIPTION:This paper aims at discovering community structure in rich me dia social networks\, through analysis of time-varying\, multi-relationa l data. Community structure represents the latent social context of user actions. It has important applications in information tasks such as sea rch and recommendation. Social media has several unique challenges. (a) In social media\, the context of user actions is constantly changing and co-evolving\; hence the social context contains time-evolving multi-dim ensional relations. (b) The social context is determined by the availabl e system features and is unique in each social media website. In this pa per we propose MetaFac (MetaGraph Factorization)\, a framework that extr acts community structures from various social contexts and interactions. Our work has three key contributions: (1) metagraph\, a novel relationa l hypergraph representation for modeling multi-relational and multi-dime nsional social data\; (2) an efficient factorization method for communit y extraction on a given metagraph\; (3) an on-line method to handle time -varying relations through incremental metagraph factorization. Extensiv e experiments on real-world social data collected from the Digg social m edia website suggest that our technique is scalable and is able to extra ct meaningful communities based on the social media contexts. We illustr ate the usefulness of our framework through prediction tasks. We outperf orm baseline methods (including aspect model and tensor analysis) by an order of magnitude. SUMMARY:MetaFac: Community Discovery via Relational Hypergraph Factorizat ion LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T173500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T172000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4975 DESCRIPTION:This paper describes and evaluates privacy-friendly methods f or extracting quasi-social networks from browser behavior on user-genera ted content sites\, for the purpose of finding good audiences for brand advertising (as opposed to click maximizing\, for example). Targeting s ocial-network neighbors resonates well with advertisers\, and on-line br owsing behavior data counterintuitively can allow the identification of good audiences anonymously. Besides being one of the first papers to ou r knowledge on data mining for on-line brand advertising\, this paper ma kes several important contributions. We introduce a framework for evalu ating brand audiences\, in analogy to predictive-modeling holdout evalua tion. We introduce methods for extracting quasi-social networks from data on visitations to social networking pages\, without collecting a ny information on the identities of the browsers or the content of the s ocial-network pages.\n\nWe introduce measures of brand proximity in the network\, and show that audiences with high brand proximity indeed show substantially higher brand affinity. Finally\, we provide evidence that the quasi-social network embeds a true social network\, which along with results from social theory offers one explanation for the increase in b rand affinity of the selected audiences. SUMMARY:Audience Selection for On-line Brand Advertising: Privacy-friendl y Social Network Targeting LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4976 DESCRIPTION:Mining discrete patterns in binary data is important for subs ampling\, compression\, and clustering. We consider rank-one binary matr ix approximations that identify the dominant patterns of the data\, whil e preserving its discrete property. A best approximation on such data ha s a minimum set of inconsistent entries\, i.e.\, mismatches between the given binary data and the approximate matrix. Due to the hardness of the problem\, previous accounts of such problems employ heuristics and the resulting approximation may be far away from the optimal one. In this pa per\, we show that the rank-one binary matrix approximation can be refor mulated as a 0-1 integer linear program (ILP). However\, the ILP formula tion is computationally expensive even for small-size matrices. We propo se a linear program (LP) relaxation\, which is shown to achieve a guaran teed approximation error bound. We further extend the proposed formulati ons using the regularization technique\, which is commonly employed to a ddress overfitting. The LP formulation is restricted to medium-size matr ices\, due to the large number of variables involved for large matrices. Interestingly\, we show that the proposed approximate formulation can b e transformed into an instance of the minimum s-t cut problem\, which ca n be solved efficiently by finding maximum flows. Our empirical study sh ows the efficiency of the proposed algorithm based on the maximum flow. Results also confirm the established theoretical bounds. SUMMARY:Mining Discrete Patterns via Binary Matrix Factorization LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T121500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T120000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4977 DESCRIPTION:The Drosophila gene expression pattern images document the sp atial and temporal dynamics of gene expression and they are valuable too ls for explicating the gene functions\, interaction\, and networks durin g Drosophila embryogenesis. To provide text-based pattern searching\, th e images in the Berkeley Drosophila Genome Project (BDGP) study are anno tated with ontology terms manually by human curators. We present a syste matic approach for automating this task\, because the number of images n eeding text descriptions is now rapidly increasing. We consider both imp roved feature representation and novel learning formulation to boost the annotation performance. For feature representation\, we adapt the bag-o f-words scheme commonly used in visual recognition problems so that the image group information in the BDGP study is retained. Moreover\, images from multiple views can be integrated naturally in this representation. To reduce the quantization error caused by the bag-of-words representat ion\, we propose an improved feature representation scheme based on the sparse learning technique. In the design of learning formulation\, we pr opose a local regularization framework that can incorporate the correlat ions among terms explicitly. We further show that the resulting optimiza tion problem admits an analytical solution. Experimental results show th at the representation based on sparse learning outperforms the bag-of-wo rds representation significantly. Results also show that incorporation o f the term-term correlations improves the annotation performance consist ently. SUMMARY:Drosophila Gene Expression Pattern Annotation Using Sparse Featur es and Term-term Interactions LOCATION:La Seine C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T172000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T170500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4978 DESCRIPTION:Topic models provide a powerful tool for analyzing large text collections by representing high dimensional data in a low dimensional subspace. Fitting a topic model given a set of training documents requir es approximate inference techniques that are computationally expensive. With today's large-scale\, constantly expanding document collections\, i t is useful to be able to infer topic distributions for new documents wi thout retraining the model.\n\nIn this paper\, we empirically evaluate t he performance of several methods for topic inference in previously unse en documents\, including methods based on Gibbs sampling\, variational i nference\, and a new method inspired by text classification. The classif ication-based inference method produces results similar to iterative inf erence methods\, but requires only a single matrix multiplication. In ad dition to these inference methods\, we present SparseLDA\, an algorithm and data structure for evaluating Gibbs sampling distributions. Empirica l results indicate that SparseLDA can be approximately 20 times faster t han traditional LDA and provide twice the speedup of previously publishe d fast sampling methods\, while also using substantially less memory. SUMMARY:Efficient Methods for Topic Model Inference on Streaming Document Collections LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T111000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4979 DESCRIPTION:We develop and evaluate an approach to causal modeling based on time series data\, collectively referred to as "grouped graphical Gra nger modeling methods." Graphical Granger modeling uses graphical modeli ng techniques on time series data and invokes the notion of "Granger cau sality" to make assertions on causality among a potentially large number of time series variables through inference on time-lagged effects. The present paper proposes a novel enhancement to the graphical Granger meth odology by developing and applying families of regression methods that a re sensitive to group information among variables\, to leverage the grou p structure present in the lagged temporal variables according to the ti me series they belong to. Additionally\, we propose a new family of algo rithms we call group boosting\, as an improved component of grouped grap hical Granger modeling over the existing regression methods with grouped variable selection in the literature (e.g group Lasso). The introductio n of group boosting methods is primarily motivated by the need to deal w ith non-linearity in the data. We perform empirical evaluation to confir m the advantage of the grouped graphical Granger methods over the standa rd (non-grouped) methods\, as well as that specific to the methods based on group boosting. This advantage is also demonstrated for the real wor ld application of gene regulatory network discovery from time-course mic roarray data. SUMMARY:Grouped Graphical Granger Modeling Methods for Temporal Causal Mo deling LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T153000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T151500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4980 DESCRIPTION:Online forums represent one type of social media that is part icularly rich for studying human behavior in information seeking and dif fusing. The way users join communities is a reflection of the changing a nd expanding of their interests toward information. In this paper\, we s tudy the patterns of user participation behavior\, and the feature facto rs that influence such behavior on different forum datasets. We find tha t\, despite the relative randomness and lesser commitment of structural relationships in online forums\, users' community joining behaviors disp lay some strong regularities. One particularly interesting observation i s that the very weak relationships between users defined by online repli es have similar diffusion curves as those of real friendships or co-auth orships. We build social selection models\, Bipartite Markov Random Fiel d (BiMRF)\, to quantitatively evaluate the prediction performance of tho se feature factors and their relationships. Using these models\, we show that some features carry supplementary information\, and the effectiven ess of different features vary in different types of forums. Moreover\, the results of BiMRF with two-star configurations suggest that the featu re of user similarity defined by frequency of communication or number of common friends is inadequate to predict grouping behavior\, but adding node-level features can improve the fit of the model. SUMMARY:User Grouping Behavior in Online Forums LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4981 DESCRIPTION:Our dynamic graph-based relational mining approach has been d eveloped to learn structural patterns in biological networks as they cha nge over time. The analysis of dynamic networks is important not only to understand life at the system-level\, but also to discover novel patter ns in other structural data. Most current graph-based data mining approa ches overlook dynamic features of biological networks\, because they are focused on only static graphs. Our approach analyzes a sequence of grap hs and discovers rules that capture the changes that occur between pairs of graphs in the sequence. These rules represent the graph rewrite rule s that the first graph must go through to be isomorphic to the second gr aph. Then\, our approach feeds the graph rewrite rules into a machine le arning system that learns general transformation rules describing the ty pes of changes that occur for a class of dynamic biological networks. Th e discovered graph-rewriting rules show how biological networks change o ver time\, and the transformation rules show the repeated patterns in th e structural changes. In this paper\, we apply our approach to biologica l networks to evaluate our approach and to understand how the biosystems change over time. We evaluate our results using coverage and prediction metrics\, and compare to biological literature. SUMMARY:Learning Patterns in the Dynamics of Biological Networks LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T151500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T145000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4982 DESCRIPTION:The most common environment in which ranking is used takes a very specific form. Users sequentially generate queries in a digital lib rary. For each query\, ranking is applied to order a set of relevant ite ms from which the user selects his favorite. This is the case when ranki ng search results for pages on the World Wide Web or for merchandize on an e-commerce site. In this work\, we present a new online ranking algor ithm\, called NoRegret KLRank. Our algorithm is designed to use "clickth rough" information as it is provided by the users to improve future rank ing decisions. More importantly\, we show that its long term average per formance will converge to the best rate achievable by any competing fixe d ranking policy selected with the benefit of hindsight. We show how to ensure that this property continues to hold as new items are added to th e set thus requiring a richer class of ranking policies. Finally\, our e mpirical results show that\, while in some context NoRegret KLRank might be considered conservative\, a greedy variant of this algorithm actuall y outperforms many popular ranking algorithms. SUMMARY:Regret-based Online Ranking for a Growing Digital Library LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T111000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4983 DESCRIPTION:We consider the problem of producing recommendations from col lective user behavior while simultaneously providing guarantees of priva cy for these users. Specifically\, we consider the Netflix Prize data se t\, and its leading algorithms\, adapted to the framework of differentia l privacy.\n\nUnlike prior privacy work concerned with cryptographically securing the computation of recommendations\, differential privacy con strains a computation in a way that precludes any inference about the un derlying records from its output. Such algorithms necessarily introduce uncertainty---i.e.\, noise---to computations\, trading accuracy for priv acy.\n\nWe find that several of the leading approaches in the Netflix Pr ize competition can be adapted to provide differential privacy\, without significantly degrading their accuracy. To adapt these algorithms\, we explicitly factor them into two parts\, an aggregation/learning phase th at can be performed with differential privacy guarantees\, and an indivi dual recommendation phase that uses the learned correlations and an indi vidual's data to provide personalized recommendations. The adaptations a re non-trivial\, and involve both careful analysis of the per-record sen sitivity of the algorithms to calibrate noise\, as well as new post-proc essing steps to mitigate the impact of this noise.\n\nWe measure the emp irical trade-off between accuracy and privacy in these adaptations\, and find that we can provide non-trivial formal privacy guarantees while st ill outperforming the Cinematch baseline Netflix provides. SUMMARY:Differentially Private Recommender Systems: Building Privacy into the Netflix Prize Contenders LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T121500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4984 DESCRIPTION:The Affinity Propagation (AP) clustering algorithm proposed b y Frey and Dueck (2007) provides an understandable\, nearly optimal summ ary of a dataset\, albeit with quadratic computational complexity. This paper\, motivated by Autonomic Computing\, extends AP to the data stream ing framework. Firstly a hierarchical strategy is used to reduce the com plexity to ${\\cal O}(N^{1+\\e})$\; the distortion loss incurred is anal yzed in relation with the dimension of the data items. Secondly\, a coup ling with a change detection test is used to cope with non-stationary da ta distribution\, and rebuild the model as needed. The presented approac h StrAP is applied to the stream of jobs submitted to the EGEE Grid\, pr oviding an understandable description of the job flow and enabling the s ystem administrator to spot online some sources of failures. SUMMARY:Toward Autonomic Grids: Analyzing the Job Flow with Affinity Stre aming LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T121500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T120000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4985 DESCRIPTION:Most of recommender systems try to find items that are most r elevant to the older choices of a given user. Here we focus on the "surp rise me" query: A user may be bored with his/her usual genre of items (e .g.\, books\, movies\, hobbies)\, and may want a recommendation that is related\, but off the beaten path\, possibly leading to a new genre of b ooks/movies/hobbies.\n\nHow would we define\, as well as automate\, this seemingly selfcontradicting request? We introduce TANGENT\, a novel rec ommendation algorithm to solve this problem. The main idea behind TANGEN T is to envision the problem as node selection on a graph\, giving high scores to nodes that are well connected to the older choices\, and at th e same time well connected to unrelated choices. The method is carefully designed to be (a) parameter-free (b) effective and (c) fast. We illust rate the benefits of TANGENT with experiments on both synthetic and real data sets. We show that TANGENT makes reasonable\, yet surprising\, hor izon-broadening recommendations. Moreover\, it is fast and scalable\, si nce it can easily use existing fast algorithms on graph node proximity. SUMMARY:TANGENT: A Novel\, 'Surprise-me' Recommendation Algorithm LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T121500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T120000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4986 DESCRIPTION:In this article\, we report our efforts in mining the informa tion encoded as clickthrough data in the server logs to evaluate and mon itor the relevance ranking quality of a commercial web search engine. We describe a metric called pSkip that aims to quantify the ranking qualit y by estimating the probability of users encountering non relevant resul ts that cost them the efforts to read and skip. A search engine with a l ower pSkip is regarded as having a better ranking quality. A key design goal of pSkip is to integrate the findings from two sets of user studies that utilize eye-tracking devices to track users browsing patterns on the search result pages\, and that use specially instrumented browsers t o actively solicit users explicit judgments on their search activities. We present the derivation of the maximum likelihood estimation of pSkip and demonstrate its efficacy in describing the user study data. The mat hematical properties of pSkip are further analyzed and compared with sev eral objective metrics as well as the cumulated gain method that uses su bjective judgments. Experimental data show that pSkip can measure aspect s of the search quality that these existing metrics are not designed or fail to address\, such as identifying the real search intents expressed in the ambiguous queries. Although effective and superior in many ways\, we also report a series of experiments that show pSkip may be influence d by system issues that are not directly related to relevance ranking\, suggesting that measurements complementary to pSkip are still needed in order to form a holistic and accurate characterization of the ranking qu ality. SUMMARY:PSkip: Estimating relevance ranking quality from web search click through data LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T173500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T172000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4987 DESCRIPTION:Learning temporal graph structures from time series data reve als important dependency relationships between current observations and histories. Most previous work focuses on learning and predicting with "s tatic" temporal graphs only. However\, in many applications such as mec hanical systems and biology systems\, the temporal dependencies might ch ange over time. In this paper\, we develop a dynamic temporal graphical models based on hidden Markov model regression and lasso-type algorithms . Our method is able to integrate two usually separate tasks\, i.e. infe rring underlying states and learning temporal graphs\, in one unified mo del. The output temporal graphs provide better understanding about compl ex systems\, i.e. how their dependency graphs evolve over time\, and ach ieve more accurate predictions. We examine our model on two synthetic d atasets as well as a real application dataset for monitoring oil-product ion equipment to capture different stages of the system\, and achieve pr omising results. SUMMARY:Learning Dynamic Temporal Graphs for Oil-production Equipment Mon itoring System LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T114500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4988 DESCRIPTION:A heterogeneous information network is an information network composed of multiple types of objects. Clustering on such a network may lead to better understanding of both hidden structures of the network a nd the individual role played by every object in each cluster. However\ , although clustering on homogeneous networks has been studied over deca des\, clustering on heterogeneous networks has not been addressed until recently.\n\nA recent study proposed a new algorithm\, RankClus\, for cl ustering on bi-typed heterogeneous networks. However\, a real-world net work may consist of more than two types\, and the interactions among mul ti-typed objects play a key role at disclosing the rich semantics that a network carries. In this paper\, we study clustering of multi-typed he terogeneous networks with a star network schema and propose a novel algo rithm\, \\NetClus\, that utilizes links across multi-typed objects to ge nerate high-quality net-clusters. An iterative enhancement method is dev eloped that leads to effective ranking-based clustering in such heteroge neous networks. Our experiments on DBLP data show that \\NetClus\\ gener ates more accurate clustering results than the baseline topic model algo rithm PLSA and the recently proposed algorithm\, RankClus. Further\, \\N etClus\\ generates informative clusters\, presenting good ranking and cl uster membership information for each attribute object in each net-clust er. SUMMARY:Ranking-Based Clustering of Heterogeneous Information Networks wi th Star Network Schema LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T172500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T171000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4989 DESCRIPTION:Active and semi-supervised learning are important techniques when labeled data are scarce. Recently a method was suggested for combin ing active learning with a semi-supervised learning algorithm that uses Gaussian fields and harmonic functions. This classifier is relational in nature: it relies on having the data presented as a partially labeled g raph (also known as a within-network learning problem). This work showed yet again that empirical risk minimization (ERM) was the best method to find the next instance to label and provided an efficient way to comput e ERM with the semi-supervised classifier. The computational problem wit h ERM is that it relies on computing the risk for all possible instances . If we could limit the candidates that should be investigated\, then we can speed up active learning considerably. In the case where the data i s graphical in nature\, we can leverage the graph structure to rapidly i dentify instances that are likely to be good candidates for labeling. Th is paper describes a novel hybrid approach of using of community finding and social network analytic centrality measures to identify good candid ates for labeling and then using ERM to find the best instance in this c andidate set. We show on real-world data that we can limit the ERM compu tations to a fraction of instances with comparable performance. SUMMARY:Using Graph-based Metrics with Empirical Risk Minimization to Spe ed Up Active Learning on Networked Data LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T115000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T113500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4990 DESCRIPTION:Merchants selling products on the Web often ask their custome rs to share their opinions and hands-on experiences on products they hav e purchased. Unfortunately\, reading through all customer reviews is dif ficult\, especially for popular items\, the number of reviews can be up to hundreds or even thousands. This makes it difficult for a potential c ustomer to read them to make an informed decision. The OpinionMiner syst em designed in this work aims to mine customer reviews of a product and extract high detailed product entities on which reviewers express their opinions. Opinion expressions are identified and opinion orientations fo r each recognized product entity are classified as positive or negative. Different from previous approaches that employed rule-based or statisti cal techniques\, we propose a novel machine learning approach built unde r the framework of lexicalized HMMs. The approach naturally integrates m ultiple important linguistic features into automatic learning. In this p aper\, we describe the architecture and main components of the system. T he evaluation of the proposed method is presented based on processing th e online product reviews from Amazon and other publicly available datase ts. SUMMARY:OpinionMiner: A Novel Machine Learning System for Web Opinion Min ing and Extraction LOCATION:Louis Armstrong A\,B\,C+D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T172000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T170500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4991 DESCRIPTION:In the past few years there has been an increasing interest\n in the analysis of process logs. Several proposed techniques\, such as w orkflow mining\, are aimed at automatically deriving the underlying work flow models. However\, current approaches only pay little attention on a n important piece of information contained in process logs: the timestam ps\, which are used to define a sequential ordering of the performed tas ks. In this work we try to overcome these limitations by explicitly incl uding time in the extracted knowledge\, thus making the temporal informa tion a first-class citizen of the analysis process. This makes it possi ble to discern between apparently identical process executions that are performed with different transition times between consecutive tasks.\n\n This paper proposes a framework for the user-interactive exploration of a condensed representation of groups of executions of a given process. T he framework is based on the use of an existing mining paradigm: Tempora lly-Annotated Sequences (TAS). These are aimed at extracting sequential patterns where each transition between two events is annotated with a ty pical transition time that emerges from input data. With the extracted T AS\, which represent sets of possible frequent executions with their typ ical transition times\, a few factorizing operators are built. These ope rators condense such executions according to possible parallel or possib le mutual exclusive executions. Lastly\, such condensed representation i s rendered to the user via the exploration graph\, namely the Temporally -Annotated Graph (TAG). The user\, the domain expert\, is allowed to exp lore the different and alternative factorizations corresponding to diffe rent interpretations of the actual executions. According to the user cho ices\, the system discards or retains certain hypotheses on actual execu tions and shows the consequent scenarios resulting from the coresponding re-aggregation of the actual data. SUMMARY:Temporal mining for interactive workflow data analysis LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T152500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T151000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4992 DESCRIPTION:Email is one of the most prevalent communication tools today\ , and solving the email overload problem is pressingly urgent. A good wa y to alleviate email overload is to automatically prioritize received me ssages according to the priorities of each user. However\, research on s tatistical learning methods for fully personalized email prioritization (PEP) has been sparse due to privacy issues\, since people are reluctant to share personal messages and importance judgments with the research c ommunity. It is therefore important to develop and evaluate PEP methods under the assumption that only limited training examples can be availabl e\, and that the system can only have the personal email data of each us er during the training and testing of the model for that user. This pape r presents the first study (to the best of our knowledge) under such an assumption. Specifically\, we focus on analysis of personal social netwo rks to capture user groups and to obtain rich features that represent th e social roles from the viewpoint of a particular user. We also develop ed a novel semi-supervised (transductive) learning algorithm that propag ates importance labels from training examples to test examples through m essage and user nodes in a personal email network. These methods togeth er enable us to obtain an enriched vector representation of each new ema il message\, which consists of both standard features of an email messag e (such as words in the title or body\, sender and receiver IDs\, etc.) and the induced social features from the sender and receivers of the mes sage. Using the enriched vector representation as the input in SVM clas sifiers to predict the importance level for each test message\, we obtai ned significant performance improvement over the baseline system (withou t induced social features) in our experiments on a multi-user data colle ction. We obtained significant performance improvement over the baseline system (without induced social features) in our experiments on a multi- user data collection: the relative error reduction in MAE was 31% in mic ro-averaging\, and 14% in macro-averaging. SUMMARY:Mining Social Networks for Personalized Email Prioritization LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T173500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T172000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4993 DESCRIPTION:The Border Gateway Protocol (BGP) is one of the fundamental c omputer communication protocols. Monitoring and mining BGP update messag es can directly reveal the health and stability of Internet routing. Her e we make two contributions: firstly we find patterns in BGP updates\, l ike self-similarity\, power-law and lognormal marginals\; secondly using these patterns\, we find anomalies. Specifically\, we develop BGP-lens\ , an automated BGP updates analysis tool\, that has three desirable prop erties: (a) It is effective\, able to identify phenomena that would othe rwise go unnoticed\, such as a peculiar `clothesline' behavior or prolon ged `spikes' that last as long as 8 hours\; (b) It is scalable\, using a lgorithms are all linear on the number of time-ticks\; and (c) It is adm in-friendly\, giving useful leads for phenomenon of interest.\n\nWe show case the capabilities of BGP-lens by identifying surprising phenomena ve rified by syadmins\, over a massive trace of BGP updates spanning 2 year s\, from the publicly available site datapository.net. SUMMARY:BGP-lens: Patterns and Anomalies in Internet Routing Updates LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T172000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T170500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4994 DESCRIPTION:Search queries are typically very short\, which means they ar e often underspecified or have senses that the user did not think of. A broad latent query aspect is a set of keywords that succinctly represen ts one particular sense\, or one particular information need\, that can aid users in reformulating such queries. We extract such broad latent a spects from query reformulations found in historical search session logs . We propose a framework under which the problem of extracting such bro ad latent aspects reduces to that of optimizing a formal objective funct ion under constraints on the total number of aspects the system can stor e\, and the number of aspects that can be shown in response to any given query. We present algorithms to find a good set of aspects\, and also to pick the best $k$ aspects matching any query. Empirical results on re al-world search engine logs show significant gains over a strong baselin e that uses single-keyword reformulations: a gain of $14\\%$ and $23\\%$ in terms of human-judged accuracy and click-through data respectively\, and around $20\\%$ in terms of consistency among aspects predicted for ``similar'' queries. This demonstrates both the importance of broad quer y aspects\, and the efficacy of our algorithms for extracting them. SUMMARY:Mining Broad Latent Query Aspects from Search Sessions LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T113500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4995 DESCRIPTION:The discovery of biclusters\, which denote groups of items th at show coherent values across a subset of all the transactions in a dat a set\, is an important type of analysis performed on real-valued data s ets in various domains\, such as biology. Several algorithms have been p roposed to find different types of biclusters in such data sets. However \, these algorithms are unable to search the space of all possible biclu sters exhaustively. Pattern mining algorithms in association analysis al so essentially produce biclusters as their result\, since the patterns c onsist of items that are supported by a subset of all the transactions. However\, a major limitation of the numerous techniques developed in ass ociation analysis is that they are only able to analyze data sets with b inary and/or categorical variables\, and their application to real-value d data sets often involves some lossy transformation such as discretizat ion or binarization of the attributes. In this paper\, we propose a nove l association analysis framework for exhaustively and efficiently mining "range support" patterns from such a data set. On one hand\, this frame work reduces the loss of information incurred by the binarization- and d iscretization-based approaches\, and on the other\, it enables the exhau stive discovery of coherent \nbiclusters. We compared the performance of our framework with two standard biclustering algorithms through the eva luation of the similarity of the cellular functions of the genes constit uting the patterns/biclusters derived by these algorithms from microarra y data. These experiments show that the real-valued patterns discovered by our framework are better enriched by small biologically interesting f unctional classes. Also\, through specific examples\, we demonstrate the ability of the RAP framework to discover functionally enriched patterns that are not found by the commonly used biclustering algorithm ISA. The source code and data sets used in this paper\, as well as the supplemen tary material\, are available at http://www.cs.umn.edu/vk/gaurav/rap. SUMMARY:An Association Analysis Approach to Biclustering LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T164000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4996 DESCRIPTION:Social media such as blogs\, Facebook\, Flickr\, etc.\, prese nts data in a network format rather than classical IID distribution. To address the interdependency among data instances\, relational learning has been proposed\, and collective inference based on network connectivi ty is adopted for prediction. However\, connections in social media are often multi-dimensional. An actor can connect to another actor for diffe rent reasons\, e.g.\, alumni\, colleagues\, living in the same city\, sh aring similar interests\, etc. Collective inference normally does not di fferentiate these connections. In this work\, we propose to extract lat ent social dimensions based on network information\, and then utilize th em as features for discriminative learning. These social dimensions de scribe diverse affiliations of actors hidden in the network\, and the di scriminative learning can automatically determine which affiliations are better aligned with the class labels. Such a scheme is preferred whe n multiple diverse relations are associated with the same network. We co nduct extensive experiments on social media data (one from a real-world blog site and the other from a popular content sharing site). Our model outperforms representative relational learning methods based on collecti ve inference\, especially when few labeled data are available. The sensi tivity of this model and its connection to existing methods are also exa mined. SUMMARY:Relational Learning via Latent Social Dimensions LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4997 DESCRIPTION:A major source of information (often the most crucial and inf ormative part) in scholarly articles from scientific journals\, proceedi ngs and books are the figures that directly provide images and other gra phical illustrations of key experimental results and other scientific co ntents. In biological articles\, a typical figure often comprises multip le panels\, accompanied by either scoped or global captioned text. Moreo ver\, the text in the caption contains important semantic entities such as protein names\, gene ontology\, tissues labels\, etc.\, relevant to t he images in the figure. Due to the avalanche of biological literature i n recent years\, and increasing popularity of various bio-imaging techni ques\, automatic retrieval and summarization of biological information f rom literature figures has emerged as a major unsolved challenge in comp utational knowledge extraction and management in the life science. We pr esent a new structured probabilistic topic model built on a realistic fi gure generation scheme to model the structurally annotated biological fi gures\, and we derive an efficient inference algorithm based on collapse d Gibbs sampling for information retrieval and visualization. The result ing program constitutes one of the key IR engines in our SLIF system tha t has recently entered the final round (4 out 70 competing systems) of t he Elsevier Grand Challenge on Knowledge Enhancement in the Life Science . Here we present various evaluations on a number of data mining tasks t o illustrate our method. SUMMARY:Structured Correspondence Topic Models for Mining Captioned Figur es in Biological Literature LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T121500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4998 DESCRIPTION:One-Class Collaborative Filtering (OCCF) is a task that natur ally emerges in recommender system settings. Typical characteristics inc lude: Only positive examples can be observed\, classes are highly imbala nced\, and the vast majority of data points are missing. The idea of int roducing weights for missing parts of a matrix has recently been shown t o help in OCCF. While existing weighting approaches mitigate the first t wo problems above\, a sparsity preserving solution that would allow to e fficiently utilize data sets with e.g.\, hundred thousands of users and items has not yet been reported. In this paper\, we study three differen t collaborative filtering frameworks: Low-rank matrix approximation\, pr obabilistic latent semantic analysis\, and maximum-margin matrix factori zation. We propose two novel algorithms for large-scale OCCF that allow to weight the unknowns. Our experimental results demonstrate their effec tiveness and efficiency on different problems\, including the Netflix Pr ize data. SUMMARY:Mind the Gaps: Weighting the Unknown in Large-Scale One-Class Col laborative Filtering LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T152000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T145500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:4999 DESCRIPTION:For a training dataset with a nonexhaustive list of classes\, i.e. some classes are not yet known and hence are not represented\, the resulting learning problem is ill-defined. In this case a sample from a missing class is incorrectly classified to one of the existing classes. For some applications the cost of misclassifying a sample could be negl igible. However\, the significance of this problem can better be acknowl edged when the potentially undesirable consequences of incorrectly class ifying a food pathogen as a nonpathogen are considered. Our research is directed towards the real-time detection of food pathogens using optical -scattering technology. Bacterial colonies consisting of the progeny of a single parent cell scatter light at 635 nm to produce unique forward-s catter signatures. These spectral signatures contain descriptive charact eristics of bacterial colonies\, which can be used to identify bacteria cultures in real time. One bottleneck that remains to be addressed is th e nonexhaustive nature of the training library. It is very difficult if not impractical to collect samples from all possible bacteria colonies a nd construct a digital library with an exhaustive set of scatter signatu res. This study deals with the real-time detection of samples from a mi ssing class and the associated problem of learning with a nonexhaustive training dataset. Our proposed method assumes a common prior for the set of all classes\, known and missing. The parameters of the prior are est imated from the samples of the known classes. This prior is then used to generate a large number of samples to simulate the space of missing cla sses. Finally a Bayesian maximum likelihood classifier is implemented us ing samples from real as well as simulated classes. Experiments performe d with samples collected for 28 bacteria subclasses favor the proposed a pproach over the state of the art. SUMMARY:Learning with a Non-exhaustive Training Dataset: Detection of Bac teria Cultures using Optical-Scattering Technology LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5000 DESCRIPTION:Collaborative filtering is the most popular approach to build recommender systems and has been successfully employed in many applicat ions. However\, it cannot make recommendations for so-called cold start users that have rated only a very small number of items. In addition\, t hese methods do not know how confident they are in their recommendations . Trust-based recommendation methods assume the additional knowledge of a trust network among users and can better deal with cold start users\, since users only need to be simply connected to the trust network. On th e other hand\, the sparsity of the user item ratings forces the trust-ba sed approach to consider ratings of indirect neighbors that are only wea kly trusted\, which may decrease its precision. In order to find a good trade-off\, we propose a random walk model combining the trust-based and the collaborative filtering approach for recommendation. The random wal k model allows us to define and to measure the confidence of a recommend ation. We performed an evaluation on the Epinions dataset and compared o ur model with existing trust-based and collaborative filtering methods. SUMMARY:TrustWalker: A Random Walk Model for Combining Trust-based and It em-based Recommendation LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T145000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T142500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5001 DESCRIPTION:Labeling text data is quite time-consuming but essential for automatic text classification. Especially\, manually creating multiple l abels for each document may become impractical when a very large amount of data is needed for training multi-label text classifiers. To minimize the human-labeling efforts\, we propose a novel multi-label active lear ning approach which can reduce the required labeled data without sacrifi cing the classification accuracy. Traditional active learning algorithms can only handle single-label problems\, that is\, each data is restrict ed to have one label. Our approach takes into account the multi-label in formation\, and select the unlabeled data which can lead to the largest reduction of the expected model loss. Specifically\, the model loss is a pproximated by the size of version space\, and the reduction rate of the size of version space is optimized with Support Vector Machines (SVM). An effective label prediction method is designed to predict possible lab els for each unlabeled data point\, and the expected loss for multi-labe l data is approximated by summing up losses on all labels according to t he most confident result of label prediction. Experiments on several rea l-world data sets (all are publicly available) demonstrate that our appr oach can obtain promising classification result with much fewer labeled data than state-of-the-art methods. SUMMARY:Effective Multi-Label Active Learning for Text Classification LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T112500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T111000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5002 DESCRIPTION:In data publishing\, anonymization techniques such as general ization and bucketization have been designed to provide privacy protecti on. In the meanwhile\, they reduce the utility of the data. It is import ant to consider the tradeoff between privacy and utility. In a paper tha t appeared in KDD 2008\, Brickell and Shmatikov proposed an evaluation m ethodology by comparing privacy gain with utility gain resulted from ano nymizing the data\, and concluded that "even modest privacy gains requir e almost complete destruction of the data-mining utility". This conclusi on seems to undermine existing work on data anonymization. In this paper \, we analyze the fundamental characteristics of privacy and utility\, a nd show that it is inappropriate to directly compare privacy with utilit y. We then observe that the privacy-utility tradeoff in data publishing is similar to the risk-return tradeoff in financial investment\, and pro pose an integrated framework for considering privacy-utility tradeoff\, borrowing concepts from the Modern Portfolio Theory for financial invest ment. Finally\, we evaluate our methodology on the Adult dataset from th e UCI machine learning repository. Our results clarify several common mi sconceptions about data utility and provide data publishers useful guide lines on choosing the right tradeoff between privacy and utility. SUMMARY:On the Tradeoff Between Privacy and Utility in Data Publishing LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T153000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T151500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5003 DESCRIPTION:Many scalable data mining tasks rely on active learning to pr ovide the most useful accurately labeled instances. However\, what if th ere are multiple labeling sources (`oracles' or `experts') with differen t but unknown reliabilities? With the recent advent of inexpensive and scalable online annotation tools\, such as Amazon's Mechanical Turk\, th e labeling process has become more vulnerable to noise - and without pri or knowledge of the accuracy of each individual labeler. This paper add resses exactly such a challenge: how to jointly learn the accuracy of la beling sources and obtain the most informative labels for the active lea rning task at hand minimizing total labeling effort. More specifically\ , we present IEThresh (Interval Estimate Threshold) as a strategy to int elligently select the expert(s) with the highest estimated labeling accu racy. IEThresh estimates a confidence interval for the reliability of e ach expert and filters out the one(s) whose estimated upper-bound confi dence interval is below a threshold - which jointly optimizes expected a ccuracy (mean) and need to better estimate the expert's accuracy (varian ce). Our framework is flexible enough to work with a wide range of diff erent noise levels and outperforms baselines such as asking all availabl e experts and random expert selection. In particular\, IEThresh achieves a given level of accuracy with less than half the queries issued by all -experts labeling and less than a third the queries required by random e xpert selection on datasets such as the UCI mushroom one. The results sh ow that our method naturally balances exploration and exploitation as it gains knowledge of which experts to rely upon\, and selects them with i ncreasing frequency. SUMMARY:Efficiently Learning the Accuracy of Labeling Sources for Selecti ve Sampling LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5004 DESCRIPTION:The availability and the accuracy of the data dictate the suc cess of a data mining application. Increasingly\, there is a need to re sort to on-line data collection to address the problem of data availabil ity. However\, participants in on-line data collection applications are naturally distrustful of the data collector as well as their peer respon dents\, resulting in inaccurate data collected as the respondents refuse to provide truthful data in fear of collusion attacks. The current ano nymity-preserving solutions for on-line data collection are unable to ad equately resist such attacks in a scalable fashion. In this paper\, we present an efficient anonymous data collection protocol for a malicious environment such as the Internet. The protocol employs cryptographic and random shuffling techniques to preserve participants' anonymity. The pr oposed method is collusion-resistant and guarantees that an attacker wil l be unable to breach an honest participant's anonymity unless she contr ols all N-1 participants. In addition\, our method is efficient and achi eved 15-42% communication overhead reduction in comparison to the prior state-of-the-art methods. SUMMARY:Collusion-Resistant Anonymous Data Collection Method LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5005 DESCRIPTION:In recent years\, the blogosphere has experienced a substanti al increase in the number of posts published daily\, forcing users to co pe with information overload. The task of guiding users through this flo od of information has thus become critical. To address this issue\, we p resent a principled approach for picking a set of posts that best covers the important stories in the blogosphere. \n\nWe define a simple and el egant notion of coverage and formalize it as a submodular optimization p roblem\, for which we can efficiently compute a near-optimal solution. I n addition\, since people have varied interests\, the ideal coverage alg orithm should incorporate user preferences in order to tailor the select ed posts to individual tastes. We define the problem of learning a perso nalized coverage function by providing an appropriate user-interaction m odel and formalizing an online learning framework for this task. We then provide a no-regret algorithm which can quickly learn a users preferen ces from limited feedback.\n\nWe evaluate our coverage and personalizati on algorithms extensively over real blog data. Results from a user study show that our simple coverage algorithm does as well as most popular bl og aggregation sites\, including Google Blog Search\, Yahoo! Buzz\, and Digg. Furthermore\, we demonstrate empirically that our algorithm can su ccessfully adapt to user preferences. We believe that our technique\, es pecially with personalization\, can dramatically reduce information over load. SUMMARY:Turning Down the Noise in the Blogosphere LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5006 DESCRIPTION:We propose two approximation algorithms for identifying\ncomm unities in dynamic social networks. Communities are\nintuitively charact erized as unusually densely knit subsets\nof a social network. This no tion becomes more problematic\nif the social interactions change over ti me. Aggregating\nsocial networks over time can radically misrepresent th e\nexisting and changing community structure. Recently\, we\nhave propos ed an optimization-based framework for modeling\ndynamic community struc ture. Also\, we have proposed\nan algorithm for finding such structure b ased on maximum\nweight bipartite matching. In this paper\, we analyze i ts performance\nguarantee for a special case where all actors can be\nob served at all times. In such instances\, we show that the\nalgorithm is a small constant factor approximation of the\noptimum. We use a similar idea to design an approximation\nalgorithm for the general case where so me individuals\nare possibly unobserved at times\, and to show that the approximation\nfactor increases twofold but remains a constant\nregardle ss of the input size. This is the first algorithm for\ninferring communi ties in dynamic networks with a provable\napproximation guarantee. We de monstrate the general algorithm\non real data sets. The results confirm the efficiency\nand effectiveness of the algorithm in identifying dynami c\ncommunities. SUMMARY:Constant-Factor Approximation Algorithms for Identifying Dynamic Communities LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T114500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5007 DESCRIPTION:We propose a novel latent factor model to accurately predict response for large scale dyadic data in the presence of features. Our ap proach is based on a model that predicts response as a multiplicative fu nction of row and column latent factors that are estimated through separ ate regressions on known row and column features. In fact\, our model pr ovides a single unified framework to address both cold and warm start sc enarios that are commonplace in practical applications like recommender systems\, online advertising\, web search\, etc. We provide scalable and accurate model fitting methods based on Iterated Conditional Mode and M onte Carlo EM algorithms. We show our model induces a stochastic process on the dyadic space with kernel (covariance) given by a polynomial func tion of features. Methods that generalize our procedure to estimate fact ors in an online fashion for dynamic applications are also considered. O ur method is illustrated on benchmark datasets and a novel content recom mendation application that arises in the context of Yahoo! Front Page. W e report significant improvements over several commonly used methods on all datasets. SUMMARY:Regression-based Latent Factor Models LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T113500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5008 DESCRIPTION:Conjoint analysis is one of the most popular market research methodologies for assessing how customers with heterogeneous preferences appraise various objective characteristics in products or services\, wh ich provides critical inputs for many marketing decisions\, e.g. optimal design of new products and target market selection. Nowadays it becomes practical in e-commercial applications to collect millions of samples q uickly. However\, the large-scale data sets make traditional conjoint an alysis coupled with sophisticated Monte Carlo simulation for parameter e stimation computationally prohibitive. In this paper\, we report a succe ssful large-scale case study of conjoint analysis on click through strea m in a real-world application at Yahoo!. We consider identifying users' heterogenous preferences from millions of click/view events and building predictive models to classify new users into segments of distinct behav ior pattern. A scalable conjoint analysis technique\, known as tensor se gmentation\, is developed by utilizing logistic tensor regression in sta ndard partworth framework for solutions. In offline analysis on the samp les collected from a random bucket of Yahoo! Front Page Today Module\, w e compare tensor segmentation against other segmentation schemes using d emographic information\, and study user preferences on article content w ithin tensor segments. Our knowledge acquired in the segmentation result s also provides assistance to editors in content management and user tar geting. The usefulness of our approach is further verified by the observ ations in a bucket test launched in Dec. 2008. SUMMARY:A Case Study of Behavior-driven Conjoint Analysis on Yahoo! Front Page Today Module LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T142500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5009 DESCRIPTION:Motivated by structural properties of the Web graph that supp ort efficient data structures for in memory adjacency queries\, we study the extent to which a large network can be compressed. Boldi and Vigna (WWW 2004)\, showed that Web graphs can be compressed down to three bi ts of storage per edge\; we study the compressibility of social networks where again adjacency queries are a fundamental primitive. To this end\ , we propose simple combinatorial formulations that encapsulate efficien t compressibility of graphs. We show that some of the problems are NP-ha rd yet admit effective heuristics\, some of which can exploit properties of social networks such as link reciprocity. Our extensive experiments show that social networks and the Web graph exhibit vastly different com pressibility characteristics. SUMMARY:On Compressing Social Networks LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T121500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5010 DESCRIPTION:The aim of data mining is to find novel and actionable insigh ts in data. However\, most algorithms typically just find a single (poss ibly non-novel/actionable) interpretation of the data even though altern atives could exist. The problem of finding an alternative to a given ori ginal clustering has received little attention in the literature. Curren t techniques (including our previous work) are unfocused/unrefined in th at they broadly attempt to find an alternative clustering but do not spe cify which properties of the original clustering should or should not be retained. In this work\, we explore a principled and flexible framework in order to find alternative clusterings of the data. The approach is p rincipled since it poses a constrained optimization problem\, so its exa ct behavior is understood. It is flexible since the user can formally sp ecify positive and negative feedback based on the existing clustering\, which ranges from which clusters to keep (or not) to making a trade-off between alternativeness and clustering quality. SUMMARY:A Principled and Flexible Framework for Finding Alternative Clust erings LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T142500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5011 DESCRIPTION:Various online social networks (OSNs) have been developed rap idly on the Internet. Researchers have analyzed different properties of such OSNs\, mainly focusing on the formation and evolution of the netwo rks as well as the information propagation over the networks. In knowle dge-sharing OSNs\, such as blogs and question answering systems\, issues on how users participate in the network and how users ``generate/contri bute'' knowledge are vital to the sustained and healthy growth of the ne tworks. However\, related discussions have not been reported in the res earch literature. \n\nIn this work\, we empirically study workloads fro m three popular knowledge-sharing OSNs\, including a blog system\, a soc ial bookmark sharing network\, and a question answering social network t o examine these properties. Our analysis consistently shows that (1) us ers' posting behavior in these networks exhibits strong daily and weekly patterns\, but the user active time in these OSNs does not follow expon ential distributions\; (2) the user posting behavior in these OSNs follo ws stretched exponential distributions instead of power-law distribution s\, indicating the influence of a small number of core users cannot domi nate the network\; (3) the distributions of user contributions on high-q uality and effort-consuming contents in these OSNs have smaller stretch factors for the stretched exponential distribution. Our study provides insights into user activity patterns and lays out an analytical foundati on for further understanding various properties of these OSNs. SUMMARY:Analyzing Patterns of User Content Generation in Online Social Ne tworks LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T120500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T115000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5012 DESCRIPTION:Stability is an important yet under-addressed issue in featur e selection from high-dimensional and small sample data. In this paper\, we show that stability of feature selection has a strong dependency on sample size. We propose a novel framework for stable feature selection w hich first identifies consensus feature groups from subsampling of train ing samples\, and then performs feature\nselection by treating each cons ensus feature group as a single entity. Experiments on both synthetic an d real-world data sets show that an algorithm developed under this frame work is effective at alleviating the problem of small sample size and le ads to more stable feature selection results and comparable or better ge neralization performance than state-of-the-art feature selection algorit hms. Synthetic data sets and algorithm source code are available at http ://www.cs.binghamton.edu/$\\sim$lyu/KDD09/. SUMMARY:Consensus Group Stable Feature Selection LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T151500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T145000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5013 DESCRIPTION:Logistic Regression is a well-known classification method tha t has been used widely in many applications of data mining\, machine lea rning\, computer vision\, and bioinformatics. Sparse logistic regression embeds feature selection in the classification framework using the L1-n orm regularization\, and is attractive in many applications involving hi gh-dimensional data. In this paper\, we propose Lassplore for solving la rge-scale sparse logistic regression. Specifically\, we formulate the pr oblem as the L1-ball constrained smooth convex optimization\, and propos e to solve the problem using the Nesterov's method\, an optimal first-or der black-box method for smooth convex optimization. One of the critical issues in the use of the Nesterov's method is the estimation of the ste p size at each of the optimization iterations. Previous approaches eithe r applies the constant step size which assumes that the Lipschitz gradie nt is known in advance\, or requires a sequence of decreasing step size which leads to slow convergence in practice. In this paper\, we propose an adaptive line search scheme which allows to tune the step size adapti vely and meanwhile guarantees the optimal convergence rate. Empirical co mparisons with several state-of-the-art algorithms demonstrate the effic iency of the proposed Lassplore algorithm for large-scale problems. SUMMARY:Large-Scale Sparse Logistic Regression LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T115000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T112500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5014 DESCRIPTION:We present an algorithm\, called the Offset Tree\, for learni ng to make decisions in situations where the payoff of only one choice i s observed\, rather than all choices. The algorithm reduces this settin g to binary classification\, allowing one to reuse any existing\, fully supervised binary classification algorithm in this partial information s etting. We show that the Offset Tree is an optimal reduction to binary classification. In particular\, it has regret at most (k-1) times the r egret of the binary classifier it uses (where k is the number of choices )\, and no reduction to binary classification can do better. This reduc tion is also computationally optimal\, both at training and test time\, requiring just O(log k) work to train on an example or make a prediction .\n\nExperiments with the Offset Tree show that it generally performs be tter than several alternative approaches. SUMMARY:The Offset Tree for Learning with Partial Labels LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T121500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5015 DESCRIPTION:Implicit user feedback\, including click-through and subseque nt browsing behavior\, is crucial for evaluating and improving the quali ty of results returned by search engines. Several recent studies \\cite{ webrevisitation\, agichtein2006\,bestbet\,foxEvaluating\,whiteVariabilit y} have used post-result browsing behavior including the sites visited\, the number of clicks\, and the dwell time on site in order to improve t he ranking of search results.\n\nIn this paper\, we first study user beh avior on sponsored search results (i.e.\, the advertisements displayed by search engines next to the ordinary\, organic results)\, and compare this behavior to that of organic results. Second\, to exploit post-resul t user behavior for better ranking of sponsored results\, we focus on id entifying patterns in user behavior and \\emph{predict} expected on-site actions in future instances. In particular\, we show how post-result be havior depends on various properties of the queries\, advertisement\, si tes\, and users\, and build a classifier using properties such as these to predict certain aspects of the user behavior. Additionally\, we devel op a generative model able to simulate observed user activity by utiliz ing a mixture of pareto distributions. We conduct experiments based on b illions of real navigation trails collected by a major search engine's b rowser toolbar. SUMMARY:Modeling and Predicting User Behavior in Sponsored Search LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5016 DESCRIPTION:Attribution of climate change to causal factors has been base d predominantly on simulations using physical climate models\, which hav e inherent limitations in describing such a complex and chaotic system. We propose an alternative\, data centric\, approach that relies on actua l measurements of climate observations and human and natural forcing fac tors. Specifically\, we develop a novel method to infer causality from s patial-temporal data\, as well as a procedure to incorporate extreme val ue modeling into our method in order to address the attribution of extre me climate events\, such as heatwaves. Our experimental results on a rea l world dataset indicate that changes in temperature are not solely acco unted for by solar radiance\, but attributed more significantly to CO2 a nd other greenhouse gases. Combined with extreme value modeling\, we als o show that there has been a significant increase in the intensity of ex treme temperatures\, and that such changes in extreme temperature are al so attributable to greenhouse gases. These preliminary results suggest t hat our approach can offer a useful alternative to the simulation-based approach to climate modeling and attribution\, and provide valuable insi ghts from a fresh perspective. SUMMARY:Spatial-temporal causal modeling for climate change attribution LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5017 DESCRIPTION:Graphs or networks can be used to model complex systems. Dete cting community structures from large network data is a classic and chal lenging task. In this paper\, we propose a novel community detection alg orithm\, which utilizes a dynamic process by contradicting the network t opology and the topology-based propinquity\, where the propinquity is a measure of the probability for a pair of nodes involved in a coherent co mmunity structure. Through several rounds of mutual reinforcement betwee n topology and propinquity\, the community structures are expected to na turally emerge. The overlapping vertices shared between communities can also be easily identified by an additional simple postprocessing. To ach ieve better efficiency\, the propinquity is incrementally calculated. We implement the algorithm on a vertex-oriented bulk synchronous parallel( BSP) model so that the mining load can be distributed on thousands of ma chines. We obtained interesting experimental results on several real net work data. SUMMARY:Parallel Community Detection on Large Networks with Propinquity D ynamics LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5018 DESCRIPTION:As clustering methods are often sensitive to parameter tuning \, obtaining stability in clustering results is an important task. In th is work\, we aim at improving clustering stability by attempting to dimi nish the influence of algorithmic inconsistencies and enhance the signal that comes from the data. We propose a mechanism that takes m clusterin gs as input and outputs $m$ clusterings of comparable quality\, which ar e in higher agreement with each other. We call our method the Clustering Agreement Process (CAP). To preserve the clustering quality\, CAP uses the same optimization procedure as used in clustering. In particular\, w e study the stability problem of randomized clustering methods (which us ually produce different results at each run). We focus on methods that a re based on inference in a combinatorial Markov Random Field (or Comraf\ , for short) of a simple topology. We instantiate CAP as inference withi n a more complex\, bipartite Comraf. We test the resulting system on fou r datasets\, three of which are medium-sized text collections\, while th e fourth is a large-scale user/movie dataset. First\, in all the four ca ses\, our system significantly improves the clustering stability measure d in terms of the macro-averaged Jaccard index. Second\, in all the four cases our system managed to significantly improve clustering quality as well\, achieving the state-of-the-art results. Third\, our system signi ficantly improves stability of consensus clustering built on top of the randomized clustering solutions. SUMMARY:Improving Clustering Stability with Combinatorial MRFs LOCATION:Miles Davis A\,B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5019 DESCRIPTION:Best Application Paper Award Runner-Up\n\n\nMotivation: Data centers are a critical component of modern IT infrastructure but are als o among the worst environmental offenders through their increasing energ y usage and the resulting large carbon footprints. Efficient management of data centers\, including power management\, networking\, and cooling infrastructure\, is hence crucial to sustainability. In the absence of a 'first-principles' approach to manage these complex components and thei r interactions\, data-driven approaches have become attractive and tenab le.\n\nResults: We present a temporal data mining solution to model and optimize performance of data center chillers\, a key component of the co oling infrastructure. It helps bridge raw\, numeric\, time-series inform ation from sensor streams toward higher level characterizations of chill er behavior\, suitable for a data center engineer. To aid in this transd uction\, temporal data streams are first encoded into a symbolic represe ntation\, next run-length encoded segments are mined to form frequent mo tifs in time series\, and finally these metrics are evaluated by their c ontributions to sustainability. A key innovation in our application is t he ability to intersperse ``don't care'' transitions (e.g.\, transients) in continuous-valued time series data\, an advantage we inherit by the application of frequent episode mining to symbolized representations of numeric time series. Our approach provides both qualitative and quantita tive characterizations of the sensor streams to the data center engineer \, to aid him in tuning chiller operating characteristics. This system i s currently being prototyped for a data center managed by HP and experim ental results from this application reveal the promise of our approach. SUMMARY:Sustainable Operation and Management of Data Center Chillers usin g Temporal Data Mining LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T153500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T152000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5020 DESCRIPTION:Classification is a core task in knowledge discovery and data mining\, and there has been substantial research effort in developing sophisticated classification models. In a parallel thread\, recent work from the NLP community suggests that for tasks such as natural language disambiguation even a simple algorithm can outperform a sophisticated o ne\, if it is provided with large quantities of high quality training da ta. In those applications\, training data occurs naturally in text corp ora\, and high quality training data sets running into billions of words have been reportedly used.\n\nWe explore how we can apply the lessons f rom the NLP community to KDD tasks. Specifically\, we investigate how to identify data sources that can yield training data at low cost and stud y whether the quantity of the automatically extracted training data can compensate for its lower quality. We carry out this investigation for th e specific task of inferring whether a search query has\ncommercial inte nt. We mine toolbar and click logs to extract queries from sites that ar e predominantly commercial (e.g.\, Amazon) and non-commercial (e.g.\, Wi kipedia). We compare the accuracy obtained using such training data aga inst manually labeled training data. Our results show that we can have large accuracy gains using automatically extracted training data at much lower cost. SUMMARY:Improving Classification Accuracy Using Automatically Extracted T raining Data LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T153000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T151500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5021 DESCRIPTION:Query result clustering has recently attracted a lot of atten tion to provide users with a succinct overview of relevant results. Howe ver\, little work has been done on organizing the query results for obje ct-level search. Object-level search result clustering is challenging be cause we need to support diverse similarity notions over object-specific features (such as the price and weight of a product) of heterogeneous d omains. To address this challenge\, we propose a hybrid subspace cluster ing algorithm called Hydra. Algorithm Hydra captures the user perception of diverse similarity notions from millions of Web pages and disambigua tes different senses using feature-based subspace locality measures. Our proposed solution\, by combining wisdom of crowds and wisdom of data\, achieves robustness and efficiency over existing approaches. We extensiv ely evaluate our proposed framework and demonstrate how to enrich user e xperiences in object-level search using a real-world product search scen arios. SUMMARY:Query Result Clustering for Object-level Search LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5022 DESCRIPTION:The explosion of user-generated content on the Web has led to new opportunities and significant challenges for companies\, that are i ncreasingly concerned about monitoring the discussion around their produ cts. Tracking such discussion on weblogs\, provides useful insight on ho w to improve products or market them more effectively. An important comp onent of such analysis is to characterize the sentiment expressed in blo gs about specific brands and products. Sentiment Analysis focuses on thi s task of automatically identifying whether a piece of text expresses a positive or negative opinion about the subject matter. Most previous wor k in this area uses prior lexical knowledge in terms of the sentiment-po larity of words. In contrast\, some recent approaches treat the task as a text classification problem\, where they learn to classify sentiment b ased only on labeled training data. In this paper\, we present a unified framework in which one can use background lexical information in terms of word-class associations\, and refine this information for specific do mains using any available training examples. Empirical results on diver se domains show that our approach performs better than using background knowledge or training data in isolation\, as well as alternative approac hes to using lexical knowledge with text classification. SUMMARY:Sentiment Analysis of Blogs by Combining Lexical Knowledge with T ext Classification LOCATION:Louis Armstrong A\,B\,C+D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T151500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T145000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5023 DESCRIPTION:Best Application Paper Award Winner\nBehavioral targeting (BT ) leverages historical user behavior to select the ads most relevant to users to display. The state-of-the-art of BT derives a linear Poisson re gression model from fine-grained user behavioral data and predicts click -through rate (CTR) from user history. We designed and implemented a hig hly scalable and efficient solution to BT using Hadoop MapReduce framewo rk. With our parallel algorithm and the resulting system\, we can build above 450 BT-category models from the entire Yahoo's user base within on e day\, the scale that one can not even imagine with prior systems. More over\, our approach has yielded 20% CTR lift over the existing productio n system by leveraging the well-grounded probabilistic model fitted from a much larger training dataset.\n\nSpecifically\, our major contributio ns include: (1) A MapReduce statistical learning algorithm and implement ation that achieve optimal data parallelism\, task parallelism\, and loa d balance in spite of the typically skewed distribution of domain data. (2) An in-place feature vector generation algorithm with linear time com plexity O(n) regardless of the granularity of sliding target window. (3) An in-memory caching scheme that significantly reduces the number of di sk IOs to make large-scale learning practical. (4) Highly efficient data structures and sparse representations of models and data to enable fast model updates. We believe that our work makes significant contributions to solving large-scale machine learning problems of industrial relevanc e in general. Finally\, we report comprehensive experimental results\, u sing industrial proprietary codebase and datasets. SUMMARY:Large-Scale Behavioral Targeting LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T171500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5024 DESCRIPTION:Social networks have become a major focus of research in rece nt years\, initially directed towards static networks but increasingly\, towards dynamic ones. In this paper\, we investigate how different pre -processing decisions and different network forces such as selection and influence affect the modeling of dynamic networks. We also present empi rical justification for some of the modeling assumptions made in dynamic network analysis (e.g.\, first-order Markovian assumption) and develop metrics to measure the alignment between links and attributes under diff erent strategies of using the historical network data. We also demonstra te the effect of attribute drift\, that is\, the importance of individua l attributes in forming links change over time. SUMMARY:Measuring the Effects of Preprocessing Decisions and Network Forc es in Dynamic Network Analysis LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T151500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T145000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5025 DESCRIPTION:To take the first step beyond keyword-based search toward ent ity-based search\, suitable token spans ("spots") on documents must be i dentified as references to real-world entities from an entity catalog. Several systems have been proposed to link spots on Web pages to entitie s in Wikipedia. They are largely based on local compatibility between t he text around the spot and textual metadata associated with the entity. Two recent systems exploit inter-label dependencies\, but in limited w ays. We propose a general collective disambiguation approach. Our prem ise is that coherent documents refer to entities from one or a few relat ed topics or domains. We give formulations for the trade-off between lo cal spot-to-entity compatibility and measures of global coherence betwee n entities. Optimizing the overall entity assignment is NP-hard. We inv estigate practical solutions based on local hill-climbing\, rounding int eger linear programs\, and pre-clustering entities followed by local opt imization within clusters. In experiments involving over a hundred manu ally-annotated Web pages and tens of thousands of spots\, our approaches significantly outperform recently-proposed algorithms. SUMMARY:Collective Annotation of Wikipedia Entities in Web Text LOCATION:Louis Armstrong & Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T171500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5026 DESCRIPTION:Malicious Web sites are a cornerstone of Internet criminal ac tivities. As a result\, there has been broad interest in developing sys tems to prevent the end user from visiting such sites. In this paper\, we describe an approach to this problem based on automated URL classific ation\, using statistical methods to discover the tell-tale lexical and host-based properties of malicious Web site URLs. These methods are abl e to learn highly predictive models by extracting and automatically anal yzing tens of thousands of features potentially indicative of suspicious URLs. The resulting classifiers obtain 95-99% accuracy\, detecting lar ge numbers of malicious Web sites from their URLs\, with only modest fal se positives. SUMMARY:Beyond Blacklists: Learning to Detect Malicious Web Sites from Su spicious URLs LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T100000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5027 DESCRIPTION:Data mining techniques use score functions to quantify how we ll a model fits a given data set. Parameters are estimated by optimisin g the fit\, as measured by the chosen score function\, and model choice is guided by the size of the scores for the different models. Since dif ferent score functions summarise the fit in different ways\, it is impor tant to choose a function which matches the objectives of the data minin g exercise. For predictive classification problems\, a wide variety of score functions exist\, including measures such as precision and recall\ , the F measure\, misclassification rate\, the area under the ROC curve (the AUC)\, and others. The first four of these require a classificatio n threshold to be chosen\, a choice which may not be easy\, or may even be impossible\, especially when the classification rule is to be applied in the future. In contrast\, the AUC does not require the specificatio n of a classification threshold\, but summarises performance over the ra nge of possible threshold choices. However\, unfortunately\, and despit e the widespread use of the AUC\, it has a previously unrecognised funda mental incoherence lying at the core of its definition. This means that using the AUC can lead to poor model choice and unecessary misclassific ations. The AUC is set in context\, its deficiency explained and the im plications illustrated - with the bottom line being that the AUC should not be used. A family of coherent alternative scores is described. The ideas are illustrated with examples from bank loans\, fraud\, face reco gnition\, and health screening. SUMMARY:Mismatched Models\, Wrong Results\, and Dreadful Decisions LOCATION:La Seine END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T100000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5028 DESCRIPTION:Data mining research has developed many algorithms for variou s analysis tasks on large and complex datasets. However\, assessing the significance of data mining results has received less attention. Analyti cal methods are rarely available\, and hence one has to use computationa lly intensive methods. Randomization approaches based on null models pro vide\, at least in principle\, a general approach that can be used to ob tain empirical p-values for various types of data mining approaches. I review some of the recent work in this area\, outlining some of the open questions and problems. SUMMARY:Randomization Methods in Data Mining LOCATION:La Seine END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T100000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5029 DESCRIPTION:Network science focuses on relationships between social enti ties. It is used widely in the social and behavioral sciences\, as well as in political science\, economics\, organizational science\, and indu strial engineering. The social network perspective has been developed o ver the last sixty years by researchers in psychology\, sociology\, and anthropology\, and more recently\, to a lesser extent\, in physics. Net work science is gaining recognition and standing in the general social and behavioral science communities as the theoretical basis for examinin g social structures. This basis has been clearly defined by many theori sts\, and the paradigm convincingly applied to important substantive pro blems. However\, the paradigm requires a new and different set of conce pts and analytic tools\, beyond those provided by standard quantitativ e (particularly\, statistical) methods. These concepts and tools are th e topics of this talk. SUMMARY:Network Science: An Introduction to Recent Statistical Approach es LOCATION:La Seine END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T190000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T173000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5030 DESCRIPTION:At KDD-2009 in Paris\, a panel on open standards and cloud co mputing addressed emerging trends for data mining applications in scienc e and industry. This report summarizes the answers from a distinguished group of thought leaders representing key software vendors in the data m ining industry.\n\nSupporting open standards and the Predictive Model Ma rkup Language (PMML) in particular\, the panel members discuss topics re garding the adoption of prevailing standards\, benefits of interoperabil ity for business users\, and the practical application of predictive mod els. We conclude with an assessment of emerging technology trends and the impact that cloud computing will have on applications as well as lic ensing models for the predictive analytics industry. SUMMARY:Panel: Open Standards and Cloud Computing LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5031 DESCRIPTION:Many organizations now devote significant fractions of their advertising/outreach budgets to online advertising\; ad-networks like Ya hoo!\, Google\, MSN have responded by constructing new kinds of economic models and perform the fundamental task of matching the most relevant a ds (selected from a large inventory) for a (query\,user) pair in a given context. Nearly all of the challenges that arise are substantially data - or model-driven (or both). Computational Advertising is a relatively n ew scientific sub-discipline at the interesection of large scale search and text analysis\, information retrieval\, statistical modeling\, machi ne learning\, optimization and microeconomics that address this match-ma king problem and provides unprecedented opportunities to data miners.\n\ nTopics covered include a comprehensive introduction to several advertis ing forms (sponsored search\, contextual adverting\, display advertising )\, revenue models (pay-per-click\, pay-per-view\, pay-per-conversion) a nd data mining challenges involved\, along with an overview of state-of- the-art techniques in the area with a detailed discussion of open proble ms. We will cover information retrieval techniques and their limitations \; data mining challenges involved in performing ad matching through cli ckstream data and challenging optimization issues that arise in display advertising. In particular\, we will cover statistical modeling techniqu es for clickstream data and explore/exploit schemes to perform online ex periments for better long-term performance using multi-armed bandit sche mes. We also discuss the close relationship of techniques used in recomm ender systems to our problem but indicate several additional issues that needs to be addressed before they become routine in computational adver tising.\n\nWe will only assume basic knowledge of statistical methods\, no prior knowledge of online advertising is required. In fact\, the firs t hour that provides an introduction to the area would be appropriate fo r all registered attendees of KDD 2009. The second half would require fa miliarity with basic concepts like regression\, probability distribution s and appreciation of issues involved in fitting statistical models to l arge scale applications. No prior knowledge of multi-armed bandits would be assumed. SUMMARY:Tutorial T1: Statistical Challenges in Computational Advertising LOCATION:Louis Armstrong A & B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5032 DESCRIPTION:While SIGKDD has traditionally enjoyed an unusually high qual ity of reviewing\, there is no doubt that publishing in SIGKDD (and othe r high quality data mining conferences) is very challenging. This is esp ecially true for young faculty\, grad students whose primary advisor is not an experienced SIGKDD author\, or people from outside the community (i.e. a biologist or mathematician who has a result that might greatly i nterest the data mining community).\n\nIn this tutorial Dr. Keogh will d emonstrate some simple ideas to enhance the probability of success in ge tting your paper published in a top data mining conference\; and after t he work is published\, getting it highly cited.\n\nThese tips and tricks are based on 12 years experience as a SIGKDD author and reviewer\, and wisdom solicited from many of the most prolific data mining researchers/ reviewers.\n\nTopics covered in the tutorial include:\n\n * Finding t he right problems to work on (80% of the battle).\n * Don't summarize \, sell! Writing abstracts that put the reviewer on your side from the s tart.\n * Getting or creating the perfect dataset.\n * Experiments that tell a story.\n * Making effective and interesting figures.\n * Getting the reviewers on your side.\n * The top-ten avoidable rea sons why papers get rejected from SIGKDD.\n * Three simple tricks to increase the number of citations to your work.\n\nWhile Dr. Keogh does n ot claim to have a “magic bullet” for publishing in SIGKDD\, his sig nificant track record of publishing in top data mining venues\, combined with extensive (and deliberately uncredited) experience in helping youn ger researchers “break-in” to SIGKDD have placed him in a unique pos ition to share useful and actionable advice.\n\nWhile writing this tutor ial Dr. Keogh\, sought and received advice from many respected data mini ng researchers\, their advice is incorporated into this tutorial. SUMMARY:Tutorial T2: How to do good research\, get it published in SIGKDD and get it cited! LOCATION:Miles Davis C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5033 DESCRIPTION:Numerous real-world datasets are in matrix form\, thus matrix algebra\, linear and multilinear\, provides important algorithmic tools for analyzing them. The main type of datasets of interest in this tutor ial are graphs. Important datasets modeled as graphs include the Interne t\, the Web\, social networks (e\,g Facebook\, LinkedIn)\, computer netw orks\, biological networks and many more.\n\nWe will discuss how we repr esent a graph as a matrix (adjacency matrix\, Laplacian) and the importa nt properties of those representations. We will then show how these prop erties are used in several important problems\, including node importanc e via random walks (Pagerank)\, community detection (METIS\, Cheeger ine quality)\, graph isomorphism and graph similarity. Important dimensional ity reduction techniques (SVD and random projections) will be discussed in the context of graph mining problems.\n\nFurthermore\, we provide a s urvey of the work on the epidemic threshold\, node proximity and center- piece subgraphs. State-of-art graph mining tools for analyzing time evol ving graphs will also be presented. Throughout the tutorial\, patterns i n static and time evolving\, weighted and unweighted real-world graphs w ill be presented.\n\nThe target audience are data mining professionals w ho wish to know the most important matrix algebra tools\, their applicat ions in large graph mining and the theory behind them.\n\nPrerequisites: Computer science background (B.Sc or equivalent)\; familiarity with und ergraduate linear algebra.\n\nDemos will be presented. SUMMARY:Tutorial T3: Large Graph-Mining: Power Tools and a Practitioner's Guide LOCATION:La Seine C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5034 DESCRIPTION:The web provides an unprecedented opportunity to evaluate ide as quickly using controlled experiments\, also called randomized experim ents\, A/B tests (and their generalizations)\, split tests\, and MultiVa riable Tests (MVT). Controlled experiments embody the best scientific de sign for establishing a causal relationship between changes and their in fluence on user-observable behavior. Data Mining and Knowledge Discovery techniques can then be used to analyze the data from such experiments. The tutorial will provide a survey and practical guide to running contro lled experiments based on the recently published survey article in the D ata Mining and Knowledge Discovery Journal\, co-authored with the two of the tutorial co-presenters (http://exp-platform.com/dmkd_survey.aspx)\, and based on the book “Always Be Testing” co-authored by the 3rd tu torial co-presenter (http://www.amazon.com/Always-Be-Testing-Complete-Op timizer/dp/0470290633). The book includes use of industry tools\, such a s Google Website Optimizer and recently ranked #2 on Amazon’s sales ra nk for computers/e-commerce books. The tutorial includes multiple real-w orld examples of actual controlled experiments (many with surprising res ults)\, a review the theory and the statistics used to plan and analyze such experiments\, and a discussion of the limitations and pitfalls that might face experimenters. Demos will be shown of some tools that suppor t controlled experiments.\n\nA video of a related talk can be found on t he videolectures website:\nhttp://videolectures.net/cikm08_kohavi_pgtce/ \n\n\nThe shorter version of the DMKD survey paper is now part of the cl ass reading for several classes at Stanford University (CS147\, CS376)\, USCD (CSE 291)\, and at the University of Washington (CSEP 510).\n\nTop ics covered include:\n\n 1. Why online experimentation using controlle d experiments is important\n 2. What you need in order to conduct a va lid experiment\n 3. Planning and Analysis of basic experiments\n 4. Benefits and limitations of experimentation\n 5. Multivariable experim ents: setup\, analysis\, interpretation\, and interactions\n 6. Archit ectures\n 7. Using online free and low-cost software services (demos)\ n 8. Challenges and advanced statistical concepts for experiments\n SUMMARY:Tutorial T4: Planning\, Running\, and Analyzing Controlled Experi ments on the Web LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5035 DESCRIPTION:In this tutorial\, we give our perspective on the keys to suc cess in application of predictive modeling to competitions like KDD Cup and real-life business intelligence projects. We argue that these two mo des of applying predictive modeling share many similarities\, but have a lso some important differences. We discuss the main success factors in p redictive modeling: domain understanding\, statistical acumen\, and appr opriate algorithmic approaches. We describe our relevant experiences in the context of three recent predictive modeling competitions where our t eam has had success (KDD Cup 2007 and 2008 and INFORMS DM challenge 2008 ) and two case studies of projects we have led at IBM Research. We also survey some of the recurring challenges and complexities in practical pr edictive modeling applications. One key issue is information leakage\, a nd we discuss its definition\, influence\, detection and avoidance. We c onsider leakage to be the silent killer of many predictive modeling proj ects\, and we demonstrate its impact on the competitions\, and discuss t he challenges in addressing it in the real-life projects. Other challeng es include framing real-life modeling objectives into predictive modelin g\, and usefully applying relational learning concepts when modeling "re al-life" complex\, relational datasets. SUMMARY:Tutorial T5: Predictive Modelling in the Wild: Success Factors in Data Mining Competitions and Real-World Applications LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5036 DESCRIPTION:As data types and data structures change to keep up with evol ving technologies and applications\, data quality problems too have evol ved and become more complex and interwoven. Data streams\, web logs\, Wi kipedias\, biomedical applications\, video streams and social networking websites generate a mind boggling variety of data types. However\, data quality mining\, the use of data mining to manage\, measure and improve data quality\, has focused mostly on addressing each category of data g litch separately as a static entity.\n\nIn this tutorial we provide a te chnical\, KDD-focused account of recent research and developments in dis covering and treating complex data anomalies in a broad range of data. I n particular\, we highlight new directions in data quality mining: (a) t he applicability and effectiveness of the methodologies for various data types such as structured\, semi-structured and stream data\, (b) the de tection of concomitant data glitches and patterns like the occurrence of outliers in data with missing values and duplicates\, or the co-occurre nce of missing values and duplicates\, (c) the design of sequential appr oaches to data quality mining\, such as workflows composed of a sequence of tasks for data quality exploration and analysis. We give an overview of past work\, introduce current research in this area including recent methods and techniques for discovering complex patterns of anomalies (e .g.\, multivariate outliers\, disguised missing values\, combination of different types of noise)\, and highlight new directions and open proble ms in data quality mining.\n\nThe tutorial includes extensive case studi es and practical examples of mining data quality problems for a variety of large datasets and data types e.g.\, relational\, XML\, data streams. We discuss illustrative examples drawn from a variety of domains like C RM\, networking\, biology\, and mobility. SUMMARY:Tutorial T6: New Directions in Data Quality Mining LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5037 DESCRIPTION:A common task in surveillance\, scientific discovery and data cleaning involves monitoring routinely collected data for anomalous eve nts. Detecting events in univariate time series data can be effectively accomplished using well-established techniques such as Box-Jenkins model s\, regression\, and statistical quality control methods. In recent year s\, however\, routinely collected data has become increasingly complex. At each time step\, the data collected can consist of multivariate vecto rs and/or be spatial in nature. For instance\, healthcare data used in d isease surveillance often consists of multivariate patient records or sp atially distributed pharmaceutical sales data. Consequently\, new event detection algorithms have been developed that not only consider temporal information but also detect spatial patterns and integrate information from multiple spatio-temporal data streams.\n\nThis tutorial will presen t algorithms for event detection\, with a focus on algorithms dealing wi th multivariate temporal and spatio-temporal data. We will introduce eve nt detection by providing a general formulation of the event detection p roblem and describing its unique challenges. In the first half of the tu torial\, we will cover algorithms for detecting events in both univariat e and multivariate temporal data. The second half will present methods f or detecting events in spatio-temporal data\, including several recently proposed multivariate approaches. SUMMARY:Tutorial T7: Event Detection LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5038 DESCRIPTION:The Web has changed our way of life and the Web 2.0 has chang ed our way of perceiving and using the Web. Data analysis is now require d in a plethora of applications that aim to enrich the experience of peo ple with the Web. We first discuss data mining for the social Web. We el aborate on social network analysis and focus on community mining\, then go over to recommendation engines and personalization. We discuss the ch allenges that emerged through the shift from the traditional Web to Web 2.0. We then focus on two issues - the need to protect Web applications from manipulation and the need to make them adaptive towards change. We first discuss manipulations/attacks in recommender systems and present c ounter-measures. We then elaborate on how changes/concept drifts can be dealt with in applications that analyze clickstream data\, monitor topic s in news and blogs\, or monitor communities and their evolution.\n\nThi s tutorial is aimed at novice researchers that have general background i n data mining and are interested in understanding the potential and chal lenges pertinent to the social Web. The participants should have a basic understanding of recommendation engines\, personalization and text mode ling for mining (vector space models). They will learn how basic techniq ues are extended and new techniques are designed for mining the Web\, es pecially the social Web. They will also learn about issues that are stil l open and require further research - research that the tutorial partici pants may decide to perform themselves.\n\nOUTLINE\nPART I: Mining the S ocial Web [Osmar Zaiane]\nPART II: Recommendations and Personalization i n the Social Web [Bamshad Mobasher]\nPART III: Dealing with Evolution in the Web [Myra Spiliopoulou]\nPART IV: Mining Web Data Streams [Olfa Nas raoui] SUMMARY:Tutorial T8: Advances in Mining the Web LOCATION:Miles Davis C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5039 DESCRIPTION:The proliferation of documents available on the Web and on co rporate intranets is driving a new wave of text mining research and appl ication. Earlier research addressed extraction of information from relat ively small collections of well-structured documents such as newswire or scientific publications. Text mining from the other corpora such as the web requires new techniques drawn from data mining\, machine learning\, NLP and IR. Text mining requires preprocessing document collections (te xt categorization\, information extraction\, term extraction)\, storage of the intermediate representations\, analysis of these intermediate rep resentations (distribution analysis\, clustering\, trend analysis\, asso ciation rules\, etc.)\, and visualization of the results. In this tutori al we will present the algorithms and methods used to build text mining systems including pre-processing techniques\, supervised leearning (e.g. \, CRF)\, entity resolution\, relationship extraction\, unsupervised lea rning and machine reading.\n\nThe tutorial will cover the state of the a rt in this rapidly growing area of research\, including recent advances in unsupervised methods for extracting facts from text and methods used for web-scale mining. We will also present several real world applicatio ns of text mining. Special emphasis will be given to lessons learned fro m years of experience in developing real world text mining systems\, inc luding how to handle informal texts such as blogs and user reviews and h ow to build scalable systems.\n\nThe instructors are Ronen Feldman and L yle Ungar. Ronen is an Associate Professor of Information Systems at the Business School of the Hebrew University in Jerusalem. He is the founde r of the ClearForest text mining corporation\, and the author of the boo k "The Text Mining Handbook" published by Cambridge University Press in 2007. Lyle is an Associate Professor of Computer and Information Science at the University of Pennsylvania.He recently returned from a sabbatica l at Google\, where he and a team built what is probably the world’s l argest named entity recognition system. SUMMARY:Tutorial T9: Real World Text Mining LOCATION:Louis Armstrong A & B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5040 DESCRIPTION:Bioinformatics is an application domain where information is naturally represented in terms of relations between heterogenous objects . Modern experimentation and data acquisition techniques allow the study of complex interactions in biological systems. This raises interesting challenges because the amount of data is huge\,some information can not be observed\, and measurements may be noisy.\n\n* Using Random Forests t o uncover bivariate interactions in high dimensional small data sets. Jo rge M. Arevalillo and Hilario Navarro\n* Identification of structurally important amino acids in proteins by graphtheoretic measures. Tammy M.K. Cheng\, Yu-En Lu and Pietro Li´o\n* Lift-based search for significant dependencies in dense data sets. Wilhelmiina Hamalainen\n* Finding Optim al Parameters for Edit-Distance Based Sequence Classification is NP-Hard . Vlado Keselj\, Haibin Liu\, Norbert Zeh\, Christian Blouin and Chris W hidden\n* Multi-Class Protein Fold Recognition using Large Margin Logic based Divide and Conquer Learning. Huma Lodhi\, Stephen Muggleton and Mi ke J.E. Sternberg\n* Protein Sequence Alignment and Intrinsic Disorder: A Substitution Matrix for an Extended Alphabet. Uros Midic\, A. Keith Du nker and Zoran Obradovic\n* Handling missing values and censored data in PCA of pharmacological matrices. Jan Ramon and Fabrizio Costa\n* Compar ing Graph-based Representations of Protein for Mining Purposes. Rabie Sa idi\, Mondher Maddouri and Engelbert M. Nguifo\n* Can we improve on the identification of Transcription Factor Binding Sites? Hugh P. Shanahan SUMMARY:W01 WORKSHOP: Statistical and Relational Learning and Mining in B ioinformatics (StReBio'09) LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5041 DESCRIPTION:This workshop will discuss the results of the KDD cup 2009. T he competition is organized around a large dataset provided by the Frenc h telecom company Orange. It is a problem of Customer Relationship Manag ement (CRM)\, a key element of modern marketing strategies. Orange offer ed the opportunity to work on a large marketing database to predict the propensity of customers to switch provider (churn)\, buy new products or services (appetency)\, or buy upgrades or add-ons proposed to them to m ake the sale more profitable (up-selling). SUMMARY:W10 WORKSHOP: KDD-Cup 2009: Fast Scoring on a Large Database (KDD cup09) LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5042 DESCRIPTION:The First ACM SIGKDD International Workshop on Knowledge Disc overy from Uncertain Data (U'09) is to discuss in depth the challenges\, opportunities and techniques on the topic of analyzing and mining uncer tain data. The theme of this workshop is to make connections among the r esearch areas of probabilistic databases\, probabilistic reasoning\, and data mining\, as well as to build bridges among the aspects of models\, data\, applications\, novel mining tasks and effective solutions. By ma king connections among different communities\, we aim at understanding e ach other in terms of scientific foundation as well as commonality and d ifferences in research methodology. SUMMARY:W11 WORKSHOP: The First ACM SIGKDD Workshop on Knowledge Discover y from Uncertain Data (U'09) LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5043 DESCRIPTION:Wide-area sensor infrastructures\, remote sensors\, RFIDs\, a nd wireless sensor networks yield massive volumes of disparate\, dynamic \, and geographically distributed data. The Sensor-KDD 2009 workshop sol icits papers that describe innovative solutions in offline data mining a nd/or real-time analysis of sensor or streaming data. Position papers th at describe the challenges and requirements for sensor data based knowle dge discovery in high-priority application domains\, as well as relevant case studies\, are particularly encouraged. SUMMARY:W02 WORKSHOP: The 3rd International Workshop on Knowledge Discove ry from Sensor Data (SensorKDD-2009) LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5044 DESCRIPTION:Computer supported communication and infrastructure are integ ral parts of modern economy. Their security is of incredible importance to a wide variety of practical domains ranging from Internet service pro viders to the banking industry and e-commerce\, from corporate networks to the intelligence community. Of interest to this workshop are novel kn owledge discovery methods addressing this field\, e.g. adaptive\, active or anticipatory approaches integrating new types of contents and protoc ols. Equally important are innovative applications demonstrating the eff ectiveness of data mining in solving real-world security problems. SUMMARY:W03 WORKSHOP: Workshop on CyberSecurity and Intelligence Informat ics (CSI-KDD) LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5045 DESCRIPTION:The goal of Visual Analytics is to derive insight from massiv e\, dynamic\, ambiguous\, and often conflicting data\; detect the expect ed and discover the unexpected\; provide timely\, defensible\, and under standable assessments\; and communicate the assessment effectively for a ction. The goal of this workshop is to raise the awareness of the KDD co mmunity for the importance of Visual Analytics and bring together resear cher from the underlying fields to bridge the gap between them—to writ e a KDD research roadmap on Visual Analytics. SUMMARY:W04 WORKSHOP: Visual Analytics and Knowledge Discovery (VAKD '09) LOCATION:Ella Fitzgerald END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5046 DESCRIPTION:Advertising\, especially online advertising\, is growing rapi dly and brings about large volumes of data along with challenging data m ining problems. Following on the success of ADKDD 2007 and 2008\, ADKDD 2009 is to be held in Paris France\, in conjunction with KDD 2009\, to p rovide a high-level international forum for the academic community and t he industry to present the state of the art of algorithms and applicatio ns of advertising.\n\nWe encourage papers that bring up and formalize ne w research problems in online advertising\, or propose novel data mining techniques for existing problems. We plan to cover (but not restricted to) the following areas: Mining for Ad Relevance and Ranking\; Audience Intelligence & User Modeling\; Content Understanding\; Search Engine Mar keting\, Optimization (SEMs\, SEOs) and Other Topics in Advertising. Acc epted papers will be achieved in ACM Digital Library and one or two pape rs will be recommended to SIGKDD Explorations.\n SUMMARY:W05 WORKSHOP: The Third International Workshop on Data Mining and Audience Intelligence for Advertising (ADKDD) LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5047 DESCRIPTION:Social networks research has come a long way since the notabl e “six-degree separation” experiment. In recent years\, social netwo rk research has advanced significantly\, thanks to the prevalence of the online social websites and the availability of a variety of offline lar ge-scale social network systems such as collaboration networks. These so cial network systems are usually characterized by the complex network st ructures and rich accompanying contextual information. Researchers are i ncreasingly interested in addressing a wide range of challenges residing in these disparate social network systems\, including identifying commo n static topological properties and dynamic properties during the format ion and evolution of these social networks\, and how contextual informat ion can help in analyzing the pertaining social networks. These issues h ave important implications on community discovery\, anomaly detection\, trend prediction and can enhance applications in multiple domains such a s information retrieval\, recommendation systems\, security and so on.\n \n\nThe third SNA-KDD '2009 aims to bring together practitioners and res earchers with a specific focus on the emerging trends and industry needs associated with the traditional Web\, the social Web\, and other forms of social networking systems. Both theoretical and experimental submiss ions are encouraged. The interesting topics include (1) data mining adva nces on the discovery and analysis of communities\, on personalization f or solitary activities (like search) and social activities (like discove ry of potential friends)\, on the analysis of user behavior in open fora (like conventional sites\, blogs and fora) and in commercial platforms (like e-auctions) and on the associated security and privacy-preservatio n challenges\; (2) social network modeling\, scalable\, customizable soc ial network infrastructure construction\, dynamic growth and evolution p atterns identification and discovery using machine learning approaches o r multi-agent based simulation. SUMMARY:W06 WORKSHOP: The 3rd Workshop on Social Network Mining and Analy sis (SNA-KDD) LOCATION:Le Jardin du Luxembourg D & E END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T083000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5048 DESCRIPTION:Human computation is a new research area that studies the pro cess of channeling the vast internet population to perform tasks or prov ide data towards solving difficult problems that no known computer algor ithms can yet solve perfectly and efficiently\, e.g. digitize books\, re cognize objects in images and songs\, translate sentences\, summarize ne ws articles\, annotate videos etc. The goal of HCOMP 2009 is to bring to gether academic and industry researchers in a stimulating discussion of existing human computation applications\, such as Games With A Purpose ( e.g. the ESP game)\, Mechanical Turk and CAPTCHAs\, and future direction s of this new subject area.\n\nIncluded in the workshop are invited talk s\, presentations\, posters\, and a demo session where participants are invited to showcase their human computation applications. SUMMARY:W07 WORKSHOP: Human Computation Workshop (HCOMP 2009) LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5049 DESCRIPTION:This workshop will present recent advances in algorithms and methods using matrix and scientific computing/applied mathematics for mo deling and analyzing massive\, high-dimensional\, and nonlinear-structur ed data. One main goal of the workshop is to bring together leading rese archers on many topic areas (e.g.\, computer scientists\, computational and applied mathematicians) to assess the state-of-the-art\, share ideas \, and form collaborations. We also wish to attract practitioners who se ek novel ideas for applications. SUMMARY:W08 WORKSHOP: Data Mining using Matrices and Tensors (DMMT'09) LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5050 DESCRIPTION:The Data Mining Case Studies Workshop and Practice Prize was established to recognize the very best data mining deployments for the y ear. Data Mining Case Studies will highlight data mining implementations that have been responsible for a significant and measurable improvement in business operations\, advanced scientific discoveries\, or provided other benefits to humanity. The best paper will be awarded the Practice Prize. Do you have an outstanding data mining application? This is a uni que opportunity to be recognized for your work. SUMMARY:W09 WORKSHOP: Data Mining Case Studies and Practice Prize (DMCS # 3) LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T145000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5051 DESCRIPTION:Web logs record the primary interaction of users with web pag es in general and search engines in particular. There are two sources f or such logs: user trails obtained from toolbars and query/click informa tion obtained from search engines. In this talk we will address the tas k of mining this rich data to improve user experience on the web. We w ill illustrate a few applications\, together with the modeling and algor ithmic challenges that stem from these applications. We will also discu ss the privacy issues that arise in this context. SUMMARY:Invited Talk - Mining Web Logs: Applications and Challenges LOCATION:Le Jardin du Luxembourg A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T145000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5052 DESCRIPTION:NASA has some of the largest and most complex data sources in the world\, with data sources ranging from the earth sciences\, space s ciences\, and massive distributed engineering data sets from commercial aircraft and spacecraft. This talk will discuss some of the issues and algorithms developed to analyze and discover patterns in these data sets . We will also provide an overview of a large research program in Integ rated Vehicle Health Management. The goal of this program is to develop advanced technologies to automatically detect\, diagnose\, predict\, an d mitigate adverse events during the flight of an aircraft. A case stud y will be presented on a recent data mining analysis performed to suppor t the Flight Readiness Review of the Space Shuttle Mission STS-119. SUMMARY:Invited Talk - Data Mining at NASA: from Theory to Applications LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5053 DESCRIPTION:This demo highlights our efforts in designing a prototype rec ommender system for use within IBM Switzerland. Goal of the system is to enable the sales and marketing teams in exploring the ever-changing wea lth of information about existing and prospective customers\, and to fac ilitate the identification of previously undiscovered sales opportunitie s. The presented system achieves its goals first by enabling the effecti ve data aggregation from diverse sources (financial data\, RSS feeds\, h ardware and software install-base\, etc)\, and then by allowing the inte ractive exploration of the customer space. Data exploration is achieved through the use of neighborhood graphs and bi-cluster formation techniqu es\, that offer both proximity and cluster navigation of the customer/pr oduct search space. The demo\, which is developed as a Web 2.0 platform\ , offers to the users the capability to interact with state-of-the-art a nalytics for recommender systems\, and also highlights the significance of exploratory data visualization techniques. SUMMARY:Demo D01 - Exploratory Recommender Systems for Sales and Marketin g LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5054 DESCRIPTION:There is an emerging focus on real-time data stream analysis on mobile/ubiquitous devices. A wide range of data stream processing app lications are targeted to run on mobile handheld devices with limited co mputational capabilities such as patient monitoring\, driver monitoring\ , providing real-time analysis and visualization for emergency calls\, o ptimization of logistics for courier pick-up and delivery etc. In this p aper\, we present the first generic toolkit for mobile data mining. The Open Mobile Miner (OMM) toolkit is easy to use\, can be deployed on a ra nge of mobile devices\, is extensible and can be customized for applicat ion specific needs. A video of the system in operation for three differe nt settings is available at: http://www.csse.monash.edu.au/~shonali/OMM/ OMMVideoDemo.asf and can be viewed using Windows Media Player. SUMMARY:Demo D02 - Open Mobile Miner: A Toolkit for Mobile Data Stream Mi ning LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5055 DESCRIPTION:Web spam\, which refers to any deliberate actions bringing to selected web pages an unjustifiable favorable relevance or importance\, is one of the major obstacles for high quality information retrieval on the web. Most of the existing web spam detection methods are supervised that require a large and representative training set of web pages. More over\, they often assume some global information such as a large web gra ph and snapshots of a large collection of web pages. However\, in many s ituations such assumptions may not hold. Recently\, we studied the probl em of online web spam detection\, and proposed the notion of spamicity t o measure how likely a page is a spam web page [9\, 7]. Spamicity is a m ore flexible and user-controllable measure than the traditional supervis ed classification methods. We developed e±cient online link spam and te rm spam detection methods using spamicity. In this paper\, we present a demonstration of OSD\, an Online Spam Detection system which can efficie ntly calculate a spamicity score online for any page on the web. SUMMARY:Demo D03 - OSD: An Online Web Spam Detection System LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5056 DESCRIPTION:This paper presents Visalix\, a Web-based interface aimed at facilitating human-computer cooperation in complex data analysis tasks. It implements an interactive visualization paradigm which assists users in matching their domain knowledge with the algorithmic power of data an alysis and mining techniques. Visalix integrates a number of Visual Inte ractive Learning components for better understanding\, easier interpreti ng complex datasets\, and training prediction models. SUMMARY:Demo D04 - Visalix: A Web Application for Visual Data Analysis an d Clustering LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5057 DESCRIPTION:With the flourishing of various Web applications\, the Intern et has become one of the most important means to access news. According to one investigation\, in the population of Internet users\, 78.5% are l ooking for news. Unfortunately\, although the Internet provides a platfo rm for easily sharing information\, it also brings a fast explosion of t he news data. It leads to the fact that people spend more and more time to digest the data. Can we design new ways to help the users quickly und erstand and explore the news data? Information retrieval is one way\, bu t it is insufficient. In this paper\, we propose a flexible topic-driven framework\, namely NewsInsight\, for news exploration. This framework i nnovatively integrates a probabilistic topic model\, graphical data anal ysis\, and natural language processing. It performs news mining at the t opic level and presents news information with topics\, entities (e.g.\, people\, organization\, and events)\, and relations derived from the new s data. Based on this framework\, we have developed a system which can h elp people to understand and explore news from multiple dimensions. The trial operation of the system has worked at Xinhua News Agency\, one of the biggest news publishers in China. Feedback from users shows that the system has achieved its primary objectives. SUMMARY:Demo D05 - A Flexible Topic-driven Framework for News Exploration LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5058 DESCRIPTION:This paper presents Model Monitor (M2)\, a Java toolkit for r obustly evaluating machine learning algorithms in the presence of changi ng data distributions. M2 provides a simple and intuitive framework in w hich users can evaluate classifiers under hypothesized shifts in distrib ution and therefore determine the best model (or models) for their data under a number of potential scenarios. Additionally\, M2 is fully integr ated with the WEKA machine learning environment\, so that a variety of c ommodity classifiers can be used if desired. SUMMARY:Demo D06 - Model Monitor: Tracking Model Performance in the Real World LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5059 DESCRIPTION:We present SHIFTR\, a system that assists users in making sen se of large scale graph data. Making sense of information represented as large graphs is a fundamental challenge in many data-intensive domains. We suggest the potential of strong synergies between the data mining\, cognitive psychology\, and HCI communities in matching powerful graph mi ning tools with insights into how people learn and interact with informa tion\, and here we present SHIFTR as one such application. SHIFTR adapts the Belief Propagation algorithm to target important sensemaking tasks such as flexibly reorganizing graph entities into multiple groups based on both positive and negative examples. SHIFTR scales linearly with the graph size through its fast algorithm\, novel mList data structure\, and externalization of graph meta data.\n\nWe demonstrate SHIFTR’s usage and benefits through real-world sensemaking scenarios using the DBLP dat aset that has almost 2 million author-publication relationships.\nA demo video of SHIFTR can be downloaded at http://www.cs.cmu.edu/~dchau/shift r/shiftr.mov. SUMMARY:Demo D07 - SHIFTR: A Fast and Scalable System for Ad Hoc Sensemak ing of Large Graphs LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5060 DESCRIPTION:We demonstrate CSAW\, a system for Curating and Searching the Annotated Web. CSAW annotates named entities and quantities in Web-scal e text corpora\, and\, where confident\, connects these annotations with entries in an entity and type catalog such as Wikipedia. The semistruct ured catalog\, together with the unstructured corpus\, forms a composite database that CSAW can then search using powerful reachability\, proxim ity and aggregation primitives. Specifically\, we can look for snippets with mentions of specific entities\, entities of a specified type\, quan tities with specified types or units\, find unions and intersections of snippet sets\, and then aggregate evidence from snippet sets into ranke d responses. Responses are not page URLs as in standard Web search\, but ranked tables where the cells can be entity references\, quantities\, o r token snippets. We will show a subset of CSAW’s capabilities\, and d escribe the beginnings of a next-generation Web search API that signific antly extends the capabilities of APIs provided by popular search engine s today. SUMMARY:Demo D08 - Curating and Searching the Annotated Web LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5061 DESCRIPTION:Expert finding\, aiming to answer the question: "Who are expe rts on topic X?"\, is becoming one of the biggest challenges for informa tion management. Much work has been conducted for expert finding. Method s based on language model\, topic model\, and random walk have been prop osed. However\, little work has studied why people want to find experts. \n\nIn this work\, we describe Expert2Bole\, a search tool that offers e xpert finding for various purposes. Specifically\, we first employ the l earning-to-rank techniques to learn a function for ranking experts. We f urther investigate a specific case of why people search experts\, i.e. B ole search\, which tries to identify best supervisors in a given field. How to learn a good ranking function for Bole search is a very challengi ng issue\, since there would be very limited or even no supervised infor mation which can be used to learn the ranking function. We propose a uni fied knowledge transfer approach which takes advantage of the expert fin ding knowledge to learn the ranking function for Bole search. A prototyp e system has been developed for expert finding and Bole search based on the proposed approach. Experiment results show the effectiveness of the proposed approach. SUMMARY:Demo D09 - Expert2Bólè: From Expert Finding to Bólè Search LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5062 DESCRIPTION:This demo presents Spam Miner\, an online system designed for real-time monitoring and characterization of spam traffic over the Inte rnet. Our system is based on high-level abstractions such as spam messag e attributes\, spam campaigns and spamming strategies. A campaign is a c luster of messages that are generated from a single message template\; c ampaign identification is a challenging problem because it has to handle spammer evolution\, while seeking for a spam similarity function that c ombines different message characteristics and for strategies that effici ently process large volumes of spams. Moreover\, spam campaigns need to be identified on-the-fly\, to allow incident response teams and security specialists to react to the threat adequately. Spam Miner addresses cam paign identification as a data clustering problem and campaigns are iden tified dynamically using a novel incremental approach based on the conce pt of Frequent Pattern Tree. Spam Miner is being used by NIC.br (Brazili an Network Information Center) and mined more than 350 million spam mess ages\, detecting meaningful clusters and patterns\, and helping the orga nization to better understand the spam problem in Brazil and how the Bra zilian Internet infrastructure is being abused by spammers.\n SUMMARY:Demo D10 - Spam Miner: A Platform for Detecting and Characterizin g Spam Campaigns LOCATION:room tbd END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T181500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T180000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5066 DESCRIPTION: SUMMARY:Opening Remarks LOCATION:La Seine END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T184500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T181500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5067 DESCRIPTION:Chair: Gregory Piatetsky-Shapiro\n* Best Paper Awards\, Wei W ang\n* Student Travel Awards\, Jure Leskovec\n* SIGKDD Dissertation Awar ds\, Bamshad Mombasher\n* Best Data Mining Case Study award from the DM CS workshop\, Gabor Melli\n* KDD Cup Winners\, David Vogel & Isabelle Gu yon\n* SIGKDD Service and Innovation Awards\, Robert Grossman\n SUMMARY:Award Presentations LOCATION:La Seine END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T190000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T184500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5068 DESCRIPTION:Dynamics of Large Networks SUMMARY:Dissertation Award Talk LOCATION:La Seine END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T193000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T190000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5069 DESCRIPTION: SUMMARY:Innovation Award Winner Talk LOCATION:La Seine END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T140000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5070 DESCRIPTION: SUMMARY:S08: Session Chair END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5071 DESCRIPTION:This session will include posters for papers that were presen ted earlier on Monday\, and that are scheduled to be presented on Wednes day. SUMMARY:Poster Session #1 LOCATION:Hôtel de Ville END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T220000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T193000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5072 DESCRIPTION:This session will include posters for papers that were presen ted earlier in the day (Tuesday). SUMMARY:Poster Session #2 END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T140000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T123000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5073 DESCRIPTION: SUMMARY:Lunch LOCATION:Foyer Rives de Seine END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T140000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T120000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5074 DESCRIPTION:* Progress Report by Gregory Piatetsky-Shapiro \n* Vision for the Future by new SIGKDD Chair SUMMARY:Lunch and SIGKDD business meeting END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T181500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T173000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5075 DESCRIPTION: SUMMARY:KDD Transfer Meeting (SIGKDD organizers only) END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T090500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5076 DESCRIPTION: SUMMARY:Plenary Keynote #2 - Session Chair END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T090500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090701T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5078 DESCRIPTION: SUMMARY:Plenary Keynote #3 - Session Chair END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T180000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T173000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5137 DESCRIPTION:The speaker will present the Europe Media Monitor (EMM) famil y of applications developed at the Joint Research Centre (JRC) of the Eu ropean Commission. EMM consists of four publicly accessible news analysi s systems (see http://press.jrc.it/overview.html):\n(1)  NewsBrief – presents the current state of affairs and detects sudden changes in rea l time\;\n(2)  MedISys – is the Medical Information System focusing specifically on health-related news\;\n(3)  NewsExplorer – allows to navigate news over time and across languages\; also gathers information about people and organisations from multilingual news in the course of t ime.\n(4)  EMM-Labs – gives access to various data visualisation and advanced text processing tools.\n\nEMM collects more than 80\,000 artic les per day from about 2\,200 online news sources (e.g. BBC\, Le Monde) in 43 different languages\, including non-Latin character set languages such as Chinese\, Arabic and Russian. EMM applications employ robust an d efficient techniques using statistics and Language Technology to clust er news articles into major news stories\, monitor the development of a story over time and across languages\, extract information about entitie s (locations\, persons and organisation) covered in the media\, and more . The major objective of EMM is to serve the needs of users in the Europ ean Commission and in European Union Member State institutions. However\ , the service is freely accessible online so that a wide range of other users benefit from the applications: EMM web sites get between one and t wo Million hits per day (approximately 30\,000 visitors per day). Additi onally\, many users are subscribed to email notifications. For more tech nical details and related research publications\, see http://langtech.jr c.it.\n\nMijail Kabadjov works at the European Commission’s Joint Res earch Centre (JRC) in Ispra (Italy)\, in the field of multilingual text summarisation and text mining. He joined the JRC from the School of Info rmatics of the University of Edinburgh (UK) where he spent two years wit h the Language Technology Group developing text mining applications for the biomedical and recruitment domains. He holds a Ph.D. in Computer Sci ence from the University of Essex (UK)\, defended with a thesis on gener al-purpose anaphora resolution. Before commencing the Ph.D.\, he worked on various projects in industry\, ranging from fraud detection in credit card transactions to customer service optimisation of a large manufactu ring firm. SUMMARY:The Europe Media Monitor family of news analysis applications LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T183000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T180000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5138 DESCRIPTION:Organizations can use data mining and advanced analytics to d ramatically improve their bottom line in three basic ways\, by (1) strea mlining a process\, (2) eliminating the bad\, or (3) improving the good. But modern organizations are so effective at their core tasks that dat a mining usually results in an iterative\, rather than transformative\, improvement. This talk will discuss the traits and culture within a bus iness unit that tend to turn data mining techincal successes into busine ss successes\, by highlighting some projects for some of America's most innovative agencies and corporations.\n\nAntonia de Medinaceli is a Seni or Business Analyst at Elder Research\, where she has been involved in a ll aspects of the data mining process over the past decade. Antonia has applied data mining technologies to a wide range of projects\, including direct marketing\, crime pattern analysis\, credit scoring\, and fraud detection. Her consulting experience is both domestic and international\ , and she is fluent in her native French and proficient in Spanish. Ant onia is experienced with most of the leading statistical software packag es\, and has taught data mining short courses both on concepts and speci fic software. She has degrees in Computer Science and Systems Engineeri ng from the University of Virginia. SUMMARY:Organizational Traits leading to High ROI for Data Mining LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T190000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090629T183000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5139 DESCRIPTION:On-line marketing spend is rapidly growing and relatively poo rly understood when compared to many of the traditional marketing channe ls. On-line marketing has characteristics that are similar to both indir ect (e.g. television\, radio) and direct (e.g. catalog) marketing. The a bility to track Web visits through “cookies” all the way through a p urchase is analogous to a catalog marketer’s use of source codes to tr ace orders back to specific catalog versions. However\, the concept of a d impressions that are viewed by an untracked audience is similar to tel evision audiences viewing a commercial. The unique nature of on-line mar keting and the available data sources leads to a need for unique applica tions for managing and optimizing the spend. This talk will present the results from three live tests of the use of data mining and constraint-b ased optimization to improve the business results for paid search\, one of the major areas of online marketing. The case studies will cover a la rge US retail bank\, a large US e-commerce site\, and a US real estate l ead aggregator.\n\nRobert Cooley holds a PhD in Computer Science from th e University of Minnesota. Currently\, Dr. Cooley is the Chief Technolog y Officer for OptiMine Software\, which specializes in analytic applicat ions for on-line marketing. Prior to OptiMine\, Dr. Cooley was a VP of T echnical Operations for KXEN Inc.\, a data mining software company. Dr. Cooley is known for his groundbreaking work in Web Mining and has over a decade of experience applying data mining to business problems. He has published numerous papers on the topics of CRM Mining\, Web Usage Mining \, and Text Mining. SUMMARY:Improving Online Marketing Performance through Data Mining and Op timization LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T180000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T173000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5140 DESCRIPTION:Datamining is expanding it's scope towards the combination of hard factual measured data with soft declared surveyed data. This new t rend enables us to understand causality of events and enables the taggin g of qualitative segmentations in the customer base. However\, both act ivities still have alot of challenges\, mainly due to the discrepancy of what people say and what they actually do. Perception versus Reality.\n \nAlain joined the France Telecom/Orange Group in Jan '04 as Director of Customer Insight. Prior to that he worked 5 years at Mobistar\, an affi liate of France Telecom in Belgium\, after being 5 years in Belgacom\, t he fixed incumbent operator in Belgium. He specialises in Segmentation w ith its customer base 'tagging'\, and in Social Network Analysis. His we ak points are his relationship with his shaver. SUMMARY: People aren't always doing what they are saying. Perception vers us Reality. LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T183000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T180000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5141 DESCRIPTION:The telecommunications industry has undergone tremendous chan ges in recent years. Advances in increased bandwidth capabilities and fu nctionality of mobile devices\, along with other factors have resulted i n more sophisticated customers\, highly competitive global markets\, new “players”\, and the emergence of new data-intensive services. All o f these changes are leading to a search for new business models\, partne rships\, and open frameworks\, which constitute a departure from the way the telecommunications industry had “traditionally” functioned. The se factors combined with a tremendous increase in data (not just in traf fic\, but also in what is available to and generated by consumers in new services) are creating huge opportunities for KDD and tremendous challe nges\, not just from a business intelligence perspective\, but also for user modeling applied to new business models and services. In this talk I will discuss how this landscape has evolved and its impact on KDD task s\, giving specific examples\, describing technical challenges and point ing to the links from research to business applications in current and f uture settings. \n\nAlejandro Jaimes is Senior Research Scientist at Tel efónica Research in Madrid where he works closely with engineering tea ms on applied research solutions on data mining and user modeling from a human-centered perspective. Dr. Jaimes obtained his Ph.D. from Columb ia University in 2003 and has worked for IBM (USA\, Japan)\, Fuji Xerox (Japan)\, Siemens (USA)\, AT&T Bell Labs (USA)\, and IDIAP Research Inst itute-EPFL (Swizterland). He holds several patents and has given numerou s invited talks and participated in panels at several international conf erences. SUMMARY: The Telecom Revolution: Where does KDD go from here? LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T190000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090630T183000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5142 DESCRIPTION:Telco operators are now realizing the tremendous value laced in the relationships among their subscribers as revealed by their callin g pattern. Social Network Analysis (SNA) goes beyond calling circle anal ysis by analyzing also how the calling circle communicates among themsel ves. These network patterns reveal who is influential and who experience s pressure from other subscribers. This information is critical in deter mining which subscribers are worth more marketing investment due to thei r potential influence on other subscribers to adopt new products. Also\, it reveals who is suddenly more vulnerable to churn even though nothing changed in their behavior ! In this talk we will talk about the notions of SNA\, a case study based on Rogers Wireless in Canada\, and how the new KXEN module KSN enabled this work.\n\nDr. Edouard Servan-Schreiber i s Assistant Director of Advanced Analytics for Europe\, Middle East and Africa. His specialty is to help businesses extract value from their dat a and insure that the sophisticated techniques of automated learning are serving business needs. Edouard has worked across industries and market s within the EMEA region. Among the topics Edouard has actively worked o n: clickstream data for customer affinity\, mobile marketing\, pricing o ptimization\, early warning in manufacturing reliability\, text mining\, and social network analysis. Edouard began practicing artificial intell igence and statistical learning models at Carnegie Mellon University for his bachelor’s degree\, before going to UC Berkeley for his PhD in Co mputer Science. After returning to his native France\, Edouard co-founde d newsfutures.com\, a technology startup offering prediction market tech nology. SUMMARY:Social network analysis for telco operators. LOCATION:Auditorium END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5144 DESCRIPTION:In 2001 the Internal Revenue Service (IRS) estimated the tax gap\, i.e. the gap between revenue owed and revenue collected\, to be ap proximately $345 billion\, of which they were able to recover only $55 b illion. It is critical for the government to reduce the tax gap and an i mportant process for doing so is audit selection. In this paper\, we pre sent a case study where data mining based methods are used to improve th e audit selection procedure at the Minnesota Department of Revenue. We d escribe the current tax audit selection process\, discuss the data from various sources as well as the issues regarding feature selection\, and explain the data mining techniques employed. On evaluation data\, data m ining methods showed an increase of 63.1% in efficiency. We also present results from actual field experiments (i.e. results of field audits per formed by auditors at the Minnesota Department of Revenue) validating th e effectiveness of data mining for audit selection. The impact of this s tudy will be a refinement of the current audit selection and tax collect ion procedures. SUMMARY:W09> Data Mining Based Tax Audit Selection: A Case Study from Min nesota Department of Revenue LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5145 DESCRIPTION:Knowledge Discovery and Data Mining techniques are now\ncommo nly used to find novel\, potentially useful\, patterns in data. Most KDD applications involve post-hoc analysis of data and are therefore mostly limited to the identification of correlations. Recent seminal work on Q uasi-Experimental Designs (Jensen\, et al.\, 2008) attempts to identify causal relationships. Controlled experiments are a standard technique us ed in multiple fields. Through randomization and proper design\, experim ents allow establishing causality scientifically\, which is why they are the gold standard in drug tests. In software development\, multiple tec hniques are used to define product requirements\; controlled experiments provide a way to assess the impact of new features on customer behavior . The Data Mining Case Studies workshop calls for describing completed i mplementations related to data mining. Over the last three years\, we bu ilt an experimentation platform system (ExP) at Microsoft\, capable of r unning and analyzing controlled experiments on web sites and services. T he goal was to accelerate innovation through trustworthy experimentation and to enable a more scientific approach to planning and prioritization of features and designs (Foley\, 2008). Along the way\, we ran many exp eriments on over a dozen Microsoft properties and had to tackle both tec hnical and cultural challenges. We previously surveyed the literature on controlled experiments and shared technical challenges (Kohavi\, et al. \, 2009). This paper focuses on problems not commonly addressed in techn ical papers: cultural challenges\, lessons\, and the ROI of running cont rolled experiments. SUMMARY:W09> Online Experimentation at Microsoft LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5146 DESCRIPTION:Original Equipment Manufacturer companies (OEMs) are facing\n more and more the challenge to increase the efficiency and reduce the co st for the service of their equipment over their lifecycle. A predictive maintenance strategy\, where the optimal time to schedule a service vis it is forecasted based on the condition of the equipment\, is often prop osed as an answer to this challenge. However\, predictive maintenance ap proaches are frequently hampered. First\, by the lack of knowledge of th e features that gives a good indication of the condition of the equipmen t. Second\, by the processing power needed to predict the future evoluti on of the features\, which in most cases is not available from the machi ne’s processor.\n\nTo overcome these problems\, we propose in this pap er to combine two approaches that are currently used separately: data mi ning and prognostics.\n\nThe proposed method consists of two steps. Firs t\, data mining and reliability estimation techniques are applied to his torical data from the field in order to optimally identify the relevant features for the condition of the equipment and the associated threshold s. Secondly\, a prediction model is fitted to the live data of the equip ment\, collected from customer’s premises\, for predicting the future evolution of these features and forecasting the time interval to the nex t maintenance action. To overcome the limited processing power of the ma chine’s processor\, this prediction part is computed on a local server which is remotely connected to the machine.\n\nThe proposed method prov ed always to retrieve\, from the datasets\, the relevant feature to be f orecasted. Validation has been done for two different industrial cases.\ n\nA first prototype of the predictive module is implemented in some cop iers and is running in live conditions\, since November 2008\, in order to check the forecast robustness. First results showed that this module offers a very good indication on when part replacement would be required \, with some level of uncertainty which decreases over time.\n\nCalculat ed business cases showed that this module will be highly beneficial for the company. Savings of approximately €4\,8 million/year worldwide are estimated. This estimate was mainly calculated by reducing labour in re active service visits and stock costs. SUMMARY:W09> A Practical Approach to Combine Data Mining and Prognostics for Improved Predictive Maintenance LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5147 DESCRIPTION:Advances in medical imaging technology have resulted in a tre mendous increase in information density for a given study. This may resu lt from increased spatial resolution\, facilitating greater anatomical d etail\, or increased contrast resolution allowing evaluation of more sub tle structures than previously possible. An increased temporal image acq uisition rate also increases the study information content. Finally\, ne w technologies enable visualization or quantification of additional tiss ue properties or contrast mechanisms.\n\nHowever\, such technological ad vances\, while potentially improving the diagnostic benefits of a study typically result in “data overload” overwhelming the ability of radi ologists to process this information. This often manifests as increased total study time\, defined as the combination of acquisition\, processin g and interpretation times\; even more critically\, the vast increase in data does not always translate to improved diagnosis/treatment selectio n. This paper describes a related series of clinically motivated data mi ning algorithms & products that extract the key\, actionable information from the vast amount of imaging data in order to ensure an improvement in patient care (via more accurate/early diagnosis) and a simultaneous r eduction in total study time.\n\nIn addition\, these applications yield decreased inter-user variability and more accurate quantitative image-ba sed measurements. While each application targets a specific clinical tas k\, they share the common methodology of transforming raw imaging data\, through knowledge-based data mining algorithms\, into clinically releva nt information. This enables users to spend less time interacting with a n image volume to extract the clinical information it contains\, while s upporting improved diagnostic accuracy by reducing the risk of accidenta l oversight of critical information. SUMMARY:W09> Mining Medical Images LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5148 DESCRIPTION:The benefit of data mining for police seems tremendous\, yet only a few limited applications are documented. This paper starts with d escribing the implementation problems of police data mining and introduc es a new approach that tries to overcome these problems in the form of a data mining system with associative memory as the main technique. This technique makes the system easier to use\, allows uncomplicated data han dling and supports many different data types. Consequently\, data prepar ation becomes easier and results contain more information. A number of D utch police forces have already been using this system for several years with over 30 users. Since the analytical process within the police is v ery knowledge-intensive\, a high level of domain expertise is essential\ , which makes it harder to find a police data miner with sufficient doma in knowledge plus technical skills in the area of databases\, statistics and data mining. The police domain also has data quality issues and a v ery diverse information need. This is why the system design tries to red uce the need for technical skills as much as possible by working with on e standard datawarehouse\, techniques that can be configured automatical ly and active user guidance. The ease of use is also ensured by integrat ing many tools and techniques from statistics\, business intelligence an d data mining into one interactive environment that does not require the analytical process to be designed beforehand. Instead\, the analysis is performed through step by step interaction. This paper discusses the be nefit of police data mining\, the design of the system\, a number of pra ctical applications\, best practices and success stories. Experiments ha ve shown a factor 20 efficiency gain\, a factor 2 prediction accuracy in crease\, a 15% drop in crime rate\, and 50% more suspect recognition. SUMMARY:W09> Data mining for intelligence led policing LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5149 DESCRIPTION:Health care research has lead to the development and availabi lity of safe and effective drugs to treat debilitating chronic diseases. The impact of this advancement has been the ever increasing costs to ob tain these medications. Prime Therapeutics\, a pharmacy benefit manager (PBM)\, strives to ameliorate some of the financial burden through the d elivery of medications via mail order pharmacy. Mail order pharmacy prov ides cost savings to both the insurer and their members\; in addition to improving overall customer satisfaction through a convenient home deliv ery system. The objective of this project was to apply various Data Mini ng techniques\, like Classification\, Clustering and Association analysi s\, to member profile and health care claims data\, to identify the link s between member characteristics and their mail order behavior. Identify ing the individual characteristics influencing mail order acceptance beh avior will help Prime Therapeutics in creating target marketing programs to improve mail order utilization. SUMMARY:W09> Identification of Independent Individualistic Predictors of Mail Order Pharmacy Prescription Utilization of Healthcare Claims LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5150 DESCRIPTION:We describe the problem of credit risk evaluation of online p ersonal loan applicants. Credit risk scoring is implemented within the d ata mining universe\, using the stochastic gradient boosting algorithm. Discussion is concentrated around the specificity of the data and proble m\, the selection of an appropriate modeling method\, determining driver s of the probability of being a good customer\, and estimation of the im pact of different predictors on this probability. The synergy of data mi ning and spatial techniques is useful for this type of problems. SUMMARY:W09> Data Mining Approach to Credit Risk Evaluation of Online Per sonal Loan Applicants LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5151 DESCRIPTION:Nowadays there is a significant amount of data mining work pe rformed outside the DBMS. This article discusses recommendations to push data mining analysis into the DBMS paying attention to data preprocessi ng (i.e. data cleaning\, summarization and transformation)\, which tends to be the most time-consuming task in data mining projects. We present a discussion of practical issues and common solutions when transforming and preparing data sets with the SQL language for data mining purposes\, based on experience from real-life projects. We then discuss general gu idelines to create variables (features) for analysis. We introduce a sim ple prototype tool that translates statistical language programs into SQ L\, focusing on data manipulation statements. Based on experience from s uccessful projects\, we present actual time performance comparisons runn ing SQL code inside the DBMS and outside running programs on a statistic al package\, obtained from data mining projects in a store\, a bank and a phone company. We highlight which steps in data mining projects are mu ch faster in the DBMS\, compared to external servers or workstations. We discuss advantages\, disadvantages and concerns from a practical standp oint based on users feedback. This article should be useful for data min ing practitioners. SUMMARY:W09> Migration of Data Mining Pre-processing into the DBMS LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5152 DESCRIPTION:This study outlines a method of determining individual custom er potential\, based solely on data present in the customer database: de scriptive information and transaction records. We define potential as th e incremental turnover that any particular company could do with their p resent customers.In order to successfully calculate this potential in a large database with multiple variables\, we propose grouping together cu stomers who “look like each other” (known as clones)\, by means of a n appropriate clustering technique: Kohonen Networks. This method is app lied to actual data sets\, and various techniques are employed to check the stability of the clusters obtained. Real potential is then determine d by means of an empirical approach: practical application to a major Fr ench retailer’s database of 5 million customers. SUMMARY:W09> Estimating Potential Customer Value using a classification t echnique to determine customer value LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5153 DESCRIPTION:Click Fraud is a challenging problem which some have called t he “Achilles Heel” of online advertising. Publishers sign up with Ad networks to display ads on their web pages. The Publishers receive a pa yout for clicks on these ads. Unfortunately this creates an incentive fo r the publisher to generate artificial clicks on ads and essentially pri nt their own money. In order to maximize scam effectiveness\, Publishers can employ sophisticated methods to cloak their attacks\, including the use of distributed networks\, hijacked browsers\, and click fraud softw are designed to mimic humans. In this paper we will describe some of the attack vectors ranging from malware to “click fraud penetrators”. W e will also describe the large-scale data mining technologies employed t o detect these programs. We conclude with some reflections on the advers arial nature of the field and some strategies for disrupting attacker ev olution. SUMMARY:W09> Clickfraud Bot Detection: High Stakes Adversarial Signal Det ection LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5154 DESCRIPTION:This paper presents a new method to analyze the link between the probabilities produced by a classification model and the variation o f its input values. The goal is to increase the predictive probability o f a given class by exploring the possible values of the input variables taken independently. The proposed method is presented in a general frame work\, and then detailed for naive Bayesian classifiers. We also demonst rate the importance of "lever variables"\, variables which can conceivab ly be acted upon to obtain specific results as represented by class prob abilities\, and consequently can be the target of specific policies. The application of the proposed method to several data sets (data proposed in the PAKDD 2007 challenge and in the KDD Cup 2009) shows that such an approach can lead to useful indicators. SUMMARY:W09> Correlation Explorations in a Classification Model LOCATION:Louis Armstrong C & D END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5164 DESCRIPTION:In this paper we consider an application of data mining techn ology to the analysis of time series data from a pilot circulating fluid ized bed (CFB) reactor. We focus on the problem of the online mass predi ction in CFB boilers. We present a framework based on switching regressi on models depending on perceived changes in the data. We analyze three a lternatives for change detection. Additionally\, a noise canceling and a state determination and windowing mechanisms are used for improving the robustness of online prediction. We validate our ideas on real data col lected from the pilot CFB boiler. SUMMARY:W02> Handling Outliers and Concept Drift in Online Mass Flow Pred iction in CFB Boilers LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5165 DESCRIPTION:It is a consensus among earth scientists that climate change will result in an increased frequency of extreme events (e.g.\, precipit ation\, snow). Streamflow forecasts and flood/drought analyses\, given t his high variability in the climatic driver (snowpack)\, are vital in th e western United States. However\, the ability to produce accurate forec asts and analyses is dependent upon the quality (accuracy) of these pred ictors (snowpack). Current snowpack datasets are based upon in-situ tele metry. Recent satellite deployments offer an alternative remote sensing data source of snowpack. The proposed research will investigate (compare ) remote sensing datasets in western U.S. watersheds in which snowpack i s the primary driver of streamflow. A comparison is made between snow wa ter equivalent (SWE) data from in-situ snowpack telemetry (SNOTEL) sites and the advanced microwave scanning radiometer – earth observing syst em (AMSR-E) aboard NASA’s Aqua satellite. Principal component techniqu es and Singular Value Decomposition are applied to determine similaritie s and differences between the datasets and investigate regional snowpack behaviors. Given the challenges (including costs\, operation and mainte nance) of deploying SNOTEL stations\, the objective of the research is t o determine if satellite based remote sensed SWE data provide a comparab le option to in-situ datasets. Watersheds investigated include the North Platte River\, the Upper Green River\, and the Upper Colorado River. Th e time period analyzed is 2003-2008\, due to the recent deployment of th e NASA Aqua satellite. Two distinct snow regions were found to behave si milarly between both datasets using principal component analysis. Singul ar Value Decomposition linked both data products with streamflow in the region and found similar behaviors among datasets. However\, only 11 of the 84 SNOTEL sites investigated correlated at a significance of 90% or greater with its corresponding AMSR-E cell. Also\, when comparing SNOTEL data with the corresponding satellite cell\, there was a consistent dif ference in the magnitude (Snow Water Equivalent) of the datasets. Finall y\, both datasets were utilized and compared in a statistically based st reamflow forecast of several gages. SUMMARY:W02> A Comparison of SNOTEL and AMSR-E Snow Water Equivalent Data sets in Western U.S. Watersheds LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5166 DESCRIPTION:To discover patterns in historical data\, climate scientists have applied various clustering methods with the goal of identifying reg ions that share some common climatological behavior. However\, past appr oaches are limited by the fact that they either consider only a single t ime period (snapshot) of multivariate data\, or they consider only a sin gle variable by using the time series data as multi-dimensional feature vector. In both cases\, potentially useful information may be lost. More over\, clusters in high-dimensional data space can be difficult to inter pret\, prompting the need for a more effective data representation. We a ddress both of these issues by employing a complex network (graph) to re present climate data\, a more intuitive model that can be used for analy sis while also having a direct mapping to the physical world for interpr etation. A cross correlation function is used to weight network edges\, thus respecting the temporal nature of the data\, and a community detect ion algorithm identifies multivariate clusters. Examining networks for c onsecutive periods allows us to study structural changes over time. We s how that communities have a climatological interpretation and that distu rbances in structure can be an indicator of climate events (or lack ther eof). Finally\, we discuss how this model can be applied for the discove ry of more complex concepts such as unknown teleconnections or the devel opment of multivariate climate indices and predictive insights. SUMMARY:W02> An Exploration of Climate Data Using Complex Networks LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5167 DESCRIPTION:Clustering is an established data mining technique for groupi ng objects based on similarity. For sensor networks one aims at grouping sensor measurements in groups of similar measurements. As sensor networ ks have limited resources in terms of available memory and energy\, a ma jor task sensor clustering is efficient computation on sensor nodes. As a dominating energy consuming task\, communication has to be reduced for a better energy efficiency. Considering memory\, one has to reduce the amount of stored information on each sensor node. For in-network cluster ing\, k-center based approaches provide k representatives out of the col lected sensor measurements. We propose EDISKCO\, an outlier aware increm ental method for efficient detection of k-center clusters. Our novel app roach is energy aware and reduces amount of required transmissions while producing high quality clustering results. In thorough experiments on s ynthetic and real world data sets\, we show that our approach outperform s a competing technique in both clustering quality and energy efficiency . Thus\, we achieve overall significantly better life times of our senso r networks. SUMMARY:W02> EDISKCO: Energy Efficient Distributed in-Sensor-Network K-ce nter Clustering with Outliers LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5168 DESCRIPTION:Monitoring biomass over large geographic regions for seasonal changes in vegetation and crop phenology is important for many applicat ions. In this paper we a present a novel clustering based change detecti on method usingMODIS NDVI time series data. We used well known EM techni que to find GMM parameters and Bayesian Information Criteria (BIC) for d etermining the number of clusters. KL Divergence measure is then used to establish the cluster correspondence across two years (2001 and 2006) t o determine changes between these two years. The changes identi ed were further analyzed for understanding phenological events. This preliminary study shows interesting relationships between key phenological events s uch as onset\, length\, end of growing seasons. SUMMARY:W02> Phenological Event Detection from Multitemporal Image Data LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5169 DESCRIPTION:Distributed PRocessing in Mobile Environments (DPRiME) is a f ramework for processing large data sets across an ad-hoc network. Develo ped to address the shortcomings of Googleís MapReduce outside of a full y-connected network\, DPRiME separates nodes on the network into a maste r and workers\; the master distributes sections of the data to available oneshop workers to process in parallel. Upon returning results to its m aster\, a worker is assigned an unfinished task. Five data mining classi fiers were implemented to process the data: decision trees\, k-means\, k -nearest neighbor\, Naive Bayes\, and artificial neural networks. Ensemb les were used so the classification tasks could be performed in parallel . This framework is well-suited for many tasks because it handles commun ications\, node movement\, node failure\, packet loss\, data partitionin g\, and result collection automatically. Therefore\, DPRiME allows users with little knowledge of networking or distributed systems to harness t he processing power of an entire network of single- and multi-hop nodes. SUMMARY:W02> Mining in a Mobile Environment LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5170 DESCRIPTION:Intra-seasonal changes in the Indian summer monsoon are gener ally characterized by its active and break (A&B) states. Existing method s for identifying the A&B states using rainfall data rely on subjective thresholds\, ignore temporal dependence in the data\, and disregard inhe rent uncertainty in their identi cation. This paper develops a method to identify intra-seasonal changes in the monsoon using a hidden Markov mo del (HMM) that allows objective classification of the monsoon states. Th e method facilitates probabilistic interpretation which is especially us eful during the transition period between the two monsoon states. The de veloped method can also be used to - (i) identify monsoon states in real time\, (ii) forecast rainfall values\, and (iii) generate synthetic dat a. Comparisons of the results from the proposed model with those from ex isting methods suggest that the new method is a promising for detecting intra-seasonal changes in the Indian summer monsoon. SUMMARY:W02> On the Identification of Intra-seasonal Changes in the India n Summer Monsoon LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5171 DESCRIPTION:In many remote sensing applications it is important to use mu ltiple sensors to be able to understand the major spatiotemporal distrib ution patterns of an observed phenomenon. A particular remote sensing ap plication addressed in this study is estimation of an important property of atmosphere\, called Aerosol Optical Depth (AOD). Remote sensing data for AOD estimation are collected from ground and satellite-based sensor s. Satellite based measurements can be used as attributes for estimation of AOD and in this way could lead to better understanding of spatiotemp oral aerosol patterns on a global scale. Ground-based AOD estimation is more accurate and is traditionally used as groundtruth information in va lidation of satellite-based AOD estimations. In contrast to this traditi onal role of ground-based sensors\, a data mining approach allows more a ctive use of ground-based measurements as labels in supervised learning of a regression model for AOD estimation from satellite measurements. Co nsidering the high operational costs of groundbased sensors\, we are stu dying a budget-cut scenario that requires a reduction in a number of gro und-based sensors. To minimize loss of information\, the objective is to retain sensors that are the most useful as a source of labeled data. Th e proposed goodness criterion for the selection is how close the accurac y of a regression model built on data from a reduced sensor set is to th e accuracy of a model built of the entire set of sensors. We developed a n iterative method that removes sensors one by one from locations where AOD can be predicted most accurately using training data from the remain ing sites. Extensive experiments on two years of globally distributed AE RONET ground-based sensor data provide strong evidence that sensors sele cted using the proposed algorithm are more informative than the competin g approaches that select sensors at random or that select sensors based on spatial diversity. SUMMARY:W02> Reduction of Ground-Based Sensor Sites for Spatio-Temporal A nalysis of Aerosols LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5172 DESCRIPTION:Current research on data stream classification mainly focuses on supervised learning\, in which a fully labeled data stream is needed for training. However\, fully labeled data streams are expensive to obt ain\, which make the supervised learning approach difficult to be applie d to real-life applications. In this paper\, we model applications\, suc h as credit fraud detection and intrusion detection\, as a one-class dat a stream classification problem. The cost of fully labeling the data str eam is reduced as users only need to provide some positive samples toget her with the unlabeled samples to the learner. Based on VFDT and POSC4.5 \, we propose our OcVFDT (One-class Very Fast Decision Tree) algorithm. Experimental study on both synthetic and real-life datasets shows that t he OcVFDT has excellent classification performance. Even 80% of the samp les in data stream are unlabeled\; the classification performance of OcV FDT is still very close to that of VFDT\, which is trained on fully labe led stream. SUMMARY:W02> OcVFDT: One-class Very Fast Decision Tree for One-class Clas sification of Data Streams LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5173 DESCRIPTION:In this paper\, we presented a frequent pattern based framewo rk for event detection in stream data\, it consists of frequent pattern discovery\, frequent pattern selection and modeling three phases: In the first phase\, a MNOE (Mining Non-Overlapping Episode) algorithm is prop osed to find the non-overlapping frequent pattern in time series. In the frequent pattern selection phase\, we proposed an EGMAMC (Episode Gener ated Memory Aggregation Markov Chain) model to help us selecting episode s which can describe stream data significantly. Then we defined feature flows to represent the instances of discovered frequent patterns and cat egorized the distribution of frequent pattern instances into three categ ories according to the spectrum of their feature flows. At last\, we pro posed a clustering algorithm EDPA (Event Detection by Pattern Aggregatio n) to aggregate strongly correlated frequent patterns together. We argue that strongly correlated frequent patterns form events and frequent pat terns in different categories can be aggregated to form different kinds of events. Experiments on real-world sensor network datasets demonstrate that the proposed MNOE algorithm is more efficient than the existing no n-overlapping episode mining algorithm and EDPA performs better when the input frequent patterns are maximal\, significant and non-overlapping. SUMMARY:W02> A Frequent Pattern Based Framework for Event Detection in Se nsor Network Stream Data LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5174 DESCRIPTION:In regression problems where the number of predictors exceeds the number of observations and the correlation between the predictors i s high\, a dimensionality reduction or a variable selection approach is demanded. In this paper we deal with a real application where we want to retrieve the physical characteristics of a combustion process from the measurements obtained with a spectroscopic sensor. This application show s up a multicollinearity problem but furthermore it is considered an ill -posed problem. Guided by this application scenario\, we propose a clust ering approach to find out homogeneous subsets of data which are embedde d in arbitrary oriented linear manifold. This model is developed under c ertain assumptions guided by a priori problem knowledge. The resulting d ivision preserves both\, the priori assumptions and the homogeneity in t he models. Thereby we break the whole problem in n subproblems improving its individual prediction accuracy versus a global solution. We show th e obtained improvements in a real application scenario related with esti mating the temperature from spectroscopic data in a remote sensing frame work. SUMMARY:W02> Supervised Clustering via Principal Component Analysis in a Retrieval Application LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5175 DESCRIPTION:The clustering validation and clustering interpretation are t he two last steps of clustering process. The validation step permits to evaluate the goodness of clustering results using some measures. Valid r esults are then generally interpreted and used in cluster analysis. The validity measures are classfied into three categories: unsupervised meas ures\, supervised measures and relative measures. Several supervised mea sures have been proposed to perform supervised evaluation such as entrop y\, purity\, F-measure\, Jaccard coefficient and Rand statistic. General ly\, these measures evaluate results according to class labels. However\ , they are not always able to distinguish interpretable clusters because most of them depends on the number of labels. This paper proposes a new supervised evaluation measure - called "homogeneity degree"- that permi ts to merge the steps of validation and interpretation. Our measure is a pplied to a real traffic data set and is used to interpret some traffic situations. Comparison with other evaluation measures shows the performa nce of our proposal. SUMMARY:W02> A Novel Measure for Validating Clustering Results Applied to Road Traffic LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5176 DESCRIPTION:Skyline queries have gained attention for supporting multicri teria analysis of large-scale datasets. While a lot of skyline algorithm s have been proposed\, most of the algorithms build upon pre-computed in dex structures that cannot generally be supported over sensor data of dy namically changing attribute values. We aim to design a scalable non-ind ex skyline computation algorithm for sensor data. More specifically\, we propose Algorithm SkyTree constructing a dynamic lattice that divides a specific region into several subregions based on a pivot point maximizi ng dominance region. Such structure enables to perform region-wise domin ance tests\, which eliminates unnecessary point-wise dominance tests. In addition\, we ensure the progressiveness that has not been supported by any non-index algorithm\, where we can identify k points maximizing the sum of dominance regions as the greedy approximation method. The k poin ts are used to reduce communication cost between sensors in computing gl obal skyline. Our evaluation results validate the efficiency of Algorith m SkyTree\, both in terms of response time and communication overhead\, over existing algorithms. SUMMARY:W02> SkyTree: Scalable Skyline Computation for Sensor Data LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5177 DESCRIPTION:In this paper\, a k-means-based clustering method applied to power quality event data is described. The data are collected by the pow er quality (PQ) monitors\, which are developed through the National PQ P roject and installed on the electricity network. The PQ monitors detect the PQ events defined as voltage sags\, swells\, and interruptions by th e IEC Standard 61000-4-30\, and collect the raw data of the event. The p roposed method aims to cope with the huge event data size and cluster th e event types so that PQ events are ultimately classified. The method he lps to manage the event data to come up with PQ assessments for the spec ific measurement points and to make comparisons of various measurement p oints in terms of PQ of the electricity network. SUMMARY:W02> Clustering of Power Quality Event Data Collected via Monitor ing Systems Installed on the Electricity Network LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5178 DESCRIPTION:The changes in rainfall and temperature patterns over India w ere detected using Mann-Kendall trend test\, Bayesian change point analy sis\, and a hidden Markov model. A regionalization method was developed to identify homogeneous regions that experience similar weather states. The regionalization helped in nding contiguous regions with strong chang e signals. The data were investigated at di erent temporal and spatial r esolution to explore the nature of changes. The study found that all Ind ia summer monsoon is stable\, but the winter or the north-east monsoon i s gradually intensifying. It also detected an abrupt drop in the winter and spring temperature over north-central India and a gradual increase i n the summer temperature over the peninsular India. Robustness of the de tected changes were evaluated using recent reanalysis dataset. SUMMARY:W02> Change Detection in Rainfall and Temperature Patterns over I ndia LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5179 DESCRIPTION:Knowledge discovery from temporal\, spatial and spatio-tempor al data is pivotal for understanding and predicting the behavior of Eart hís ecosystem model. An important influence leaving its impact on the e cosystem is the global climate system. In this paper\, the Earth Science data that we have analyzed consists of daily global air temperature and precipitation measurements\, aggregated from heterogeneous sensors for fifty years (1950-1999). The enormous amount of data that is available f or analysis requires employment of data mining techniques for discoverin g interesting patterns\, detecting significant changes and extracting me aningful insights from the data. Our work considers the problem of detec ting anomalous (abnormal or unexpected) behavior in the global climate s ystem\, discovering teleconnection patterns and providing consequential insights to the analysts. SUMMARY:W02> Anomaly Detection and Spatio-Temporal Analysis of Global Cli mate System LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5180 DESCRIPTION:Climate modeling and analysis of climate change have largely been based on forward simulation with physical models. We propose here a data centric approach to climate study based solely on the actual obser ved data. This novel approach utilizes a variety of relevant statistical modeling and machine learning techniques such as spatial-temporal causa l modeling and extreme value modeling\, and suggests multiple future res earch directions. We will describe preliminary results using data for No rth America from CRU\, NOAA\, NASA\, NCDC\, and CDIAC\, as well as certa in technical challenges encountered. It is hoped that this alternative p erspective will help uncover new insights\, improve aspects of simulatio n models with known uncertainties\, and provide a useful complementary a pproach to climate study. SUMMARY:W02> A data modeling approach to climate change attribution LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5181 DESCRIPTION:Sensor networking is a paradigm getting familiar in space mis sions and services. The talk will provide a panoramic view of examples o f challenging missions related to earth environment and to space explora tion\, where sensor knowledge discovery techniques might become instrume ntal to fulfill mission objectives. ESA missions such as GMES (Global Mo nitoring for Environment and Security) and the series of possible Mars e xploration missions will be presented and put in context with the topic of the workshop. SUMMARY:W02> Space Missions & Sensor Networking: Challenging Scenarios LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5182 DESCRIPTION:In this talk\, we tackle a fundamental problem that arises wh en using sensors to monitor the ecological condition of rivers and lakes \, the network of pipes that bring water to our taps\, or the activities of an elderly individual when sitting on a chair: Where should we place the sensors in order to make effective and robust predictions? Such sen sing problems are typically NP-hard\, and in the past\, heuristics witho ut theoretical guarantees about the solution quality have often been use d. In this talk\, we present algorithms which efficiently find provably near-optimal solutions to large\, complex sensing problems. Our algorith ms are based on the key insight that many important sensing problems exh ibit submodularity\, an intuitive diminishing returns property: Adding a sensor helps more the fewer sensors we have placed so far. In addition to identifying most informative locations for placing sensors\, our alg orithms can handle settings\, where sensor nodes need to be able to reli ably communicate over lossy links\, where mobile robots are used for col lecting data or where solutions need to be robust against adversaries an d sensor failures. We present results applying our algorithms to several real-world sensing tasks\, including environmental monitoring using rob otic sensors\, activity recognition using a built sensing chair\, and a sensor placement competition. We conclude with drawing an interesting co nnection between sensor placement for water monitoring and addressing th e challenges of information overload on the web. As examples of this co nnection\, we address the problem of selecting blogs to read in order to learn about the biggest stories discussed on the web\, and personalizin g content to turn down the noise in the blogosphere. SUMMARY:W02> How Optimized Environmental Sensing Helps Address Informatio n Overload on the Web LOCATION:St Germain des Prés A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T091000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5183 DESCRIPTION: SUMMARY:W07> Herd It: Designing A Social Game to Tag Music LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T093500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T092500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5184 DESCRIPTION: SUMMARY:W07> KissKissBan: A Competitive Human Computation Game for Image Annotation LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T095500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T094000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5185 DESCRIPTION: SUMMARY:W07> Community-based Game Design: Experiments on Social Games for Commonsense Data Collection LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5186 DESCRIPTION: SUMMARY:W07> A Demonstration Of Human Computation Using The Phrase Detect ives Annotation Game LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5187 DESCRIPTION: SUMMARY:W07> Picture This: Preferences for Image Search LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5188 DESCRIPTION: SUMMARY:W07> Page Hunt: Using Human Computation Games to Improve Web Sear ch LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5189 DESCRIPTION: SUMMARY:W07> TurKit: Tools for Iterative Tasks on Mechanical Turk LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5190 DESCRIPTION: SUMMARY:W07> Search War: A Game for Improving Web Search LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5191 DESCRIPTION: SUMMARY:W07> Magic Bullet: A Dual-Purpose Computer Game LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5192 DESCRIPTION: SUMMARY:W07> Seaweed: A Web Application for End Users to Design Economic Games LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5193 DESCRIPTION: SUMMARY:W07> Thumbs-Up: A Game for Playing to Rank Search Results LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5194 DESCRIPTION: SUMMARY:W07> Games for Games: Manipulating Contexts in Human Computation Games LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5195 DESCRIPTION: SUMMARY:W07> From Active Towards InterActive Learning: Using Consideratio n Information to Improve Labeling Correctness LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5196 DESCRIPTION: SUMMARY:W07> TagCaptcha: Annotating images with CAPTCHAs LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5197 DESCRIPTION: SUMMARY:W07> CAPTCHA-based Image Labeling on the Soylent Grid\, Peter Fay monville LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5198 DESCRIPTION: SUMMARY:W07> Designing Crowdsourcing Community for the Enterprise LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5199 DESCRIPTION: SUMMARY:W07> A Reputation System for Selling Human Computation LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T111000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5200 DESCRIPTION: SUMMARY:W07> The Role of Game Theory in Human Computation Systems LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T113000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T111500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5201 DESCRIPTION: SUMMARY:W07> On Formal Models for Social Verification LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T115000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T113500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5202 DESCRIPTION: SUMMARY:W07> Efficient Human Computation: the Distributed Labeling Proble m LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T121000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T115500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5203 DESCRIPTION: SUMMARY:W07> Financial Incentives and the ìPerformance of Crowdsî LOCATION:La Sorbonne A\, B & C END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T092500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5204 DESCRIPTION: SUMMARY:W05> Ad Quality on TV: Predicting Television Audience Retention LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T095000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T092500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5205 DESCRIPTION: SUMMARY:W05> Handling Missing Values in GPS Surveys Using Survival Analys is LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T105500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5206 DESCRIPTION: SUMMARY:W05> A Markov Chain Model for Integrating Behavioral Targeting in to Contextual Ads LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T112000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T105500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5207 DESCRIPTION: SUMMARY:W05> Probabilistic Latent Semantic User Segmentation for Behavior al Targeted Advertising LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T114500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T112000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5208 DESCRIPTION: SUMMARY:W05> Argo: Intelligent Advertising by Mining a User's Interest fr om His Photo Collections LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T121000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T114500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5209 DESCRIPTION: SUMMARY:W05> Scalable Clustering and Keyword Suggestion for Online Advert isements LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T121000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5210 DESCRIPTION: SUMMARY:W05> Inferring Local Synonyms for Improving Keyword Suggestion in an On-line Advertisement System LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T150000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5211 DESCRIPTION: SUMMARY:W05> Invited Talk: Brand Advertising\, On-line Audiences\, and So cial Media LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T152500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T150000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5212 DESCRIPTION: SUMMARY:W05> Data-Driven Text Features for Sponsored Search Click Predict ion LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T162500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5213 DESCRIPTION: SUMMARY:W05> Revenue Optimization with Relevance Constraint in Sponsored Search LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T162500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5214 DESCRIPTION: SUMMARY:W05> Pricing Guidance in Ad Sale Negotiations: The PrintAds Examp le LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T171500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T165000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5215 DESCRIPTION: SUMMARY:W05> Online Allocation of Display Advertisements Subject to Advan ced Sales Contracts LOCATION:Miles Davis B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5216 DESCRIPTION:Phishing emails usually contain a message from a credible loo king source requesting a user to click a link to a website where she/he is asked to enter a password or other confidential information. Most phi shing emails aim at withdrawing money from financial institutions or get ting access to private information. Phishing has increased enormously ov er the last years and is a serious threat to global security and economy . There are a number of possible countermeasures to phishing. These rang e from communication-oriented approaches like authentication protocols o ver blacklisting to content-based filtering approaches. We argue that t he first two approaches are currently not broadly implemented or exhibit deficits. Therefore content-based phishing filters are necessary and wi dely used to increase communication security. A number of features are e xtracted capturing the content and structural properties of the email. S ubsequently a statistical classifier is trained using these features on a training set of emails labeled as ham (legitimate)\, spam or phishing. This classifier may then be applied to an email stream to estimate the classes of new incoming emails. AntiPhish is a specific targeted resear ch project funded under Framework Program 6 by the European Union. It is aims at developing improved anti-phishing technologies that help to pro tect and secure the global email communication infrastructure. The proje ct on the one hand developed the filter methodology in a test laboratory setting\, but on the other hand implemented this technology in real wor ld settings\, to be used to filter all email traffic online in real time . In this talk we summarize our experience with phishing filtering with benchmark data and in addition with different real-life email streams. First we describe a number of novel features that are particularly well- suited to identify phishing emails [1]. These include statistical models for the low-dimensional descriptions of email topics\, sequential analy sis of email text and external links\, the detection of embedded logos a s well as indicators for hidden salting [2]. Hidden salting is the inten tional addition or distortion of content not perceivable by the reader. For empirical evaluation we have obtained a large realistic corpus of em ails pre-labeled as spam\, phishing\, and ham (legitimate). In experimen ts with benchmark data our methods outperform other published approaches for classifying phishing emails. The second part of the talk describes the application of these approaches to real-life email streams. On the one hand we investigate how we can identify new phishing emails arriving from a honeypot system. This allows to spot new types of phishing mails . Subsequently the characteristics of these new phishing emails can be u sed to update client-based phishing filters. A second experiment investi gates the capabilities of the AntiPhish system when monitoring emails in an ISP framework. It turns out that active learning approaches are very efficient to maintain and improve filtering accuracy. We discuss the im plications of these results for the practical application of this approa ch in the workflow of an email provider. Finally we describe a strategy how the filters may be updated and adapted to new types of phishing. SUMMARY:W03> Invited Talk "AntiPhish ñ Lessons Learnt" LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5217 DESCRIPTION:Combining Incremental Hidden Markov Model and Adaboost Algori thm for Anomaly Intrusion Detection\,"Chen Yo-Shu\, Chen Yi-Ming"\,"Trad itional Hidden Markov Model (HMM) has been successfully applied to anoma ly intrusion detection. Incremental HMM (IHMM) further improves the trai ning time of HMM. However\, both HMM and IHMM still have the problem of high false positive rate. In this paper\, we propose an Adaboost-IHMM to combine IHMM and adaboost for anomaly intrusion detection. As adaboost firstly uses many IHMMs to collectively classify samples then decides th e results of samplesí classifications\, the Adaboost-IHMM can improve t he accurate rate of classifications. Experimental results with Stide dat asets show that the proposed method can significantly improve the false positive rate by 70% without decreasing detection rate. Besides\, we als o propose a method to adjust the normal profile for avoiding erroneous d etection caused by changes of normal behavior. We perform with experimen ts with realistic datasets extracted from the use of popular browsers. C ompared with traditional HMM method\, our method can improve the trainin g time by 90% to build a new normal profile. SUMMARY:W03> Combining Incremental Hidden Markov Model and Adaboost Algor ithm for Anomaly Intrusion Detection LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T113000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5218 DESCRIPTION:In network traffic monitoring\, and more particularly in the realm of threat intelligence\, the problem of attack attribution"" refer s to the process of effectively attributing new attack events to (un)-kn own phenomena\, based on some evidence or traces left on one or several monitoring platforms. Real-world attack phenomena are often largely dist ributed on the Internet\, or can sometimes evolve quite rapidly. This ma kes them inherently complex and thus difficult to analyze. In general\, an analyst must consider many different attack features (or criteria) in order to decide about the plausible root cause of a given attack\, or t o attribute it to some given phenomenon. In this paper\, we introduce a global analysis method to address this problem in a systematic way. Our approach is based on a novel combination of a knowledge discovery techni que with a fuzzy inference system\, which somehow mimics the reasoning o f an expert by implementing a multi-criteria decision-making process bui lt on top of the previously extracted knowledge. By applying this method on attack traces\, we are able to identify largescale attack phenomena with a high degree of confidence. In most cases\, the observed phenomena can be attributed to so-called zombie armies - or botnets\, i.e. groups of compromised machines controlled remotely by a same entity. By means of experiments with real-world attack traces\, we show how this method c an effectively help us to perform a behavioral analysis of those zombie armies from a long-term\, strategic viewpoint SUMMARY:W03> Addressing the Attack Attribution Problem using Knowledge Di scovery and Multi-criteria Fuzzy Decision-Making LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T120000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T113000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5219 DESCRIPTION:Commercial anti-virus software are unable to provide protecti on against newly launched (a.k.a ìzero-dayî) malware. In this paper\, we propose a novel malware detection technique which is based on the ana lysis of byte-level file content. The novelty of our approach\, compared with existing content based mining schemes\, is that it does not memori ze specific byte-sequences or strings appearing in the actual file conte nt. Our technique is non-signature based and therefore has the potential to detect previously unknown and zero-day malware. We compute a wide ra nge of statistical and information-theoretic features in a block-wise ma nner to quantify the byte-level file content. We leverage standard data mining algorithms to classify the file content of every block as normal or potentially malicious. Finally\, we correlate the block-wise classifi cation results of a given file to categorize it as benign or malware. Si nce the proposed scheme operates at the byte-level file content\; theref ore\, it does not require any a priori information about the filetype. W e have tested our proposed technique using a benign dataset comprising o f six different filetypes ó DOC\, EXE\, JPG\, MP3\, PDF and ZIP and a m alware dataset comprising of six different malware types ó backdoor\, t rojan\, virus\, worm\, constructor and miscellaneous. We also perform a comparison with existing data mining based malware detection techniques. The results of our experiments show that the proposed nonsignature base d technique surpasses the existing techniques and achieves more than 90% detection accuracy. SUMMARY:W03> Malware Detection Using Statistical Analysis of Byte-Level F ile Content LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T120000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5220 DESCRIPTION:In adversarial systems\, the performance of a classifier de- creases after it is deployed\, as the adversary learns to defeat it. Rec ently\, adversarial data mining was introduced as a solution to this\, w here the classification problem is viewed as a game mechanism between an adversary and an intelligent and adaptive classifier. Over the last yea rs\, phishing fraud through malicious email messages has been a serious threat that affects global security and economy\, where traditional spam Filtering techniques have shown to be ineffective. In this domain\, usi ng dynamic games of incomplete information\, a game theoretic data minin g framework is proposed in order to build an adversary aware classifier for phishing fraud detection. To build the classifier\, an online versio n of the Weighted Margin Support Vector Machines with a game theoretic p rior knowledge function is proposed. In this pa- per\, a new content-bas ed feature extraction technique for phishing filtering is described. Exp eriments show that the proposed classifier is highly competitive compare d with previously proposed online classification algorithms in this adve rsarial environment\, and promising results where obtained using traditi onal machine learning techniques over extracted features. SUMMARY:W03> Online Phishing Classification Using Adversarial Data Mining and Signaling Games LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T144000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5221 DESCRIPTION:Data is a critical resource in numerous organizations. One of the challenging problems facing these organizations today is to ensure that only authorized individuals have address to data. Data also has to be protected from malicious corruption. Much of the early work on data s ecurity focused on multilevel secure data management systems where users have different clearance levels and data has different sensitivity leve ls and access to data is governed by the security policies. There were m any efforts on securing relational\, distributed and object oriented dat abases. More recently\, several aspects of data security are being inves tigated including data confidentiality\, integrity\, trust and privacy. Furthermore\, securing data warehouses\, semantic web\, as well as apply ing data mining for solving security problems are getting a lot of atten tion This presentation will review the developments in data security and integrity as well as discuss directions for further research and develo pment. In particular\, policy management for the semantic web\, assured information sharing\, privacy preserving data mining and novel ways to b uild secure data management systems will be discussed. SUMMARY:W03> Invited Talk: Data Security and Integrity: Developments and Directions LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T150500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T144000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5222 DESCRIPTION:While millions of dollars have been invested in information t echnologies to improve intelligence information sharing among law enforc ement agencies at the Federal\, Tribal\, State and Local levels\, there remains a hesitation to share information between agencies. This lack of coordination hinders the ability to prevent and respond to crime and te rrorism. Work to date has not produced solutions nor widely accepted par adigms for understanding the problem. Therefore\, to enhance the current intelligence information sharing services between government entities\, in this interdisciplinary research\, we have identified three major are as of influence\; Technical\, Social\, and Legal. Furthermore\, we have developed a preliminary model and theory of intelligence information sha ring through a literature review\, experience and interviews with practi tioners in the field. This model and theory should serve as a basic conc eptual framework for further academic work and lead to further investiga tion and clarification of the identified factors and the degree of impac t they exert on the system so that actionable solutions can be identifie d and implemented. SUMMARY:W03> Towards Trusted Intelligence Information Sharing LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T153000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T150500 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5223 DESCRIPTION:Intelligence and law enforcement force make use of terrorist and criminal social networks to support their investigations such as ide ntifying suspects\, terrorist or criminal subgroups\, and their communic ation patterns. Social networks are valuable resources but it is not eas y to obtain information to create a complete terrorist or criminal socia l network. Missing information in a terrorist or criminal social network always diminish the effectiveness of investigation. An individual agenc y only has a partial terrorist or criminal social network due to their l imited information sources. Sharing and integration of social networks b etween different agencies increase the effectiveness of social network a nalysis. Unfortunately\, information sharing is usually forbidden due to the concern of privacy preservation. In this paper\, we introduce the K NN algorithm for subgraph generation and a mechanism to integrate the ge neralized information to conduct social network analysis. Generalized in formation such as lengths of the shortest paths\, number of nodes on the boundary\, and the total number of nodes is constructed for each genera lized subgraphs. By utilizing the generalized information shared from ot her sources\, an estimation of distance between nodes is developed to co mpute closeness centrality. Two experiments have been conducted with ran dom graphs and the Global Salafi Jihad terrorist social network. The res ult shows that the proposed technique improves the accuracy of closeness centrality measures substantially while protecting the sensitive data. SUMMARY:W03> Social Networks Integration and Privacy Preservation using S ubgraph Generalization LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T163000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5224 DESCRIPTION:The United States and its Allied Forces have had tremendous s uccess in combat operations. This includes combat in Germany\, Japan and more recently in Iraq and Afghanistan. However not all of our stabiliza tion and reconstruction operations (SARO) have been as successful. Recen tly several studies have been carried out on SARO by National Defense Un iversity as well as for the Army Science and Technology. One of the majo r conclusions is that we need to plan for SARO while we are planning for combat. That is\, we cannot start planning for SARO after the enemy reg ime has fallen. In addition\, the studies have shown that security\, pow er and jobs are key ingredients for success during SARO. It is important to give positions to some of the power players from the fallen regime p rovided they are trustworthy. It is critical that investments are made t o stimulate the local economies. The studies have also analyzed the vari ous technologies that are needed for successfully carrying out SARO whic h includes sensors\, robotics and information management. In this projec t we will focus on the information management component for SARO. As sta ted in the work by the Naval Postgraduate School\, we need to determine the social\, political and economic relationships between the local comm unities as well as determine who the important people are. This work has also identified the 5Ws (Who\, When\, What\, Where and Why) and the (H) . To address the key technical challenges for SARO\, we are defining a L ife cycle for SARO and subsequently developing a Temporal Geosocial Serv ice Oriented Architecture System (TGSSOA) that utilizes Temporal Geosoci al Semantic Web (TGS-SW) technologies for managing this lifecycle. We ar e developing techniques for representing temporal geosocial information and relationships\, integrating such information and relationships\, que rying such information and relationships and finally reasoning about suc h information and relationships so that the commander can answer questio ns related to the 5Ws and H. To our knowledge we believe that this is th e first attempt to develop. SUMMARY:W03> Design of a Temporal Geosocial Semantic Web for Military Sta bilization and Reconstruction Operation LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T170000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T163000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5225 DESCRIPTION:developed\, used\, and criticized frequently in the recent pa st. This paper examines several of the more common criticisms and analyz es some factors that bear on whether the criticisms are valid and/or can be overcome by appropriate design and use of the data mining applicatio n. SUMMARY:W03> On the Efficacy of Data Mining for Security Applications LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T170000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5226 DESCRIPTION:It is believed that public companies should have put lots of efforts and resources in designing and implementing effective security p olicy in their daily information processing and management against poten tial cyber attacks. A company web server accessible by the general publi c and attackers is usually a common entry point for cyber attacks. This paper studies and reports the security problems in web servers of public companies. We applied several commonly used tools and systems to collec t information from publicly accessible web servers of selected public co mpanies\, and studied some known security aspects in those public compan ies. Our findings will provide an insight to the effectiveness of web se rvers in public companies against cyber attacks. This paper also propose s a risk analysis tool for cyber attacks\, which is known as pyramid ris k analysis tool. SUMMARY:W03> A Study of Online Service and Information Exposure of Public Companies LOCATION:Miles Davis A END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5227 DESCRIPTION:Random Forests (RF) is an ensemble method which has become wi dely accepted within the machine learning and bioinformatics communities in the last few years. Its predictive strength\, along with some of the ingredients ---rich in information--- provided by the output\, has made RF an efficient Data Mining tool for discovering patterns in data. In t his paper we review the learning mechanism of RF within the classificati on setting and apply it to uncover bivariate interactions\, carrying on useful information about an outcome\, in high dimensional low sample dat a. We propose a divide and conquer search strategy in the variable space that benefits from the ranking of variable importances of RF at a first stage\, along with the out of bag error rate (oob) of the ensemble at a second stage. The procedure combines both elements in order to capture difficult to uncover patterns in these type of data. We will show the pe rformance of our procedure in some synthetic scenarios and will give a r eal application to a microarray data set in order to illustrate how it w orks. SUMMARY:W01> Using Random Forests to uncover bivariate interactions in hi gh dimensional small data sets LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153739Z UID:5228 DESCRIPTION: SUMMARY:W01> Identification of structurally important amino acids in prot eins by graph-theoretic measures LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5229 DESCRIPTION:Dependency analysis is an important but computationally deman ding problem in all empirical science. It is especially problematic in b ioinformatics\, where data sets are often high dimensional\, dense and/o r strongly correlated. As a solution\, we introduce a new algorithm whic h searches the most significant association rules expressing positive de pendencies. The algorithm uses several effective pruning principles\, wh ich enable search without any minimum frequency thresholds. According to our initial experiments\, the algorithm suits especially well for typic al biological and medical data sets. SUMMARY:W01> Lift-based search for significant dependencies in dense data sets LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5230 DESCRIPTION:Parametric edit distance based classification has been applie d to two significant problems in the bioinformatics area: biological seq uence analysis (DNA\, RNA\, protein)\, and semantic relationship extract ion from biomedical scientific literature. This method is based on the edit distance measure on sequences\, with parametric costs for matching\ , mis-matching\, inserts\, and deletes of letters. We present a proof th at finding optimal parameter values for such classification based on tra ining data is an NP-hard problem\, which is an important claim to justif y the use of heuristic methods for determining the best parameter values . SUMMARY:W01> Finding Optimal Parameters for Edit Distance Based Sequence Classification is NP-Hard LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T123000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T103000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5231 DESCRIPTION:Inductive Logic Programming (ILP) systems have been successfu lly applied to solve complex biological problem by viewing them as binar y classification tasks. It remains an open question how an accurate solu tion to a multi-class problem can be obtained by using a logic based lea rning method. In this paper we present a novel logic based approach to s olve complex and challenging multi-class classification problems in bioi nformatics by focusing on a particular task\, namely protein fold recogn ition. Our technique is based on the use of large margin kernel-based me thods in conjunction with first order rules induced by an ILP system. Th e proposed approach learns a multi-class classifier by using a divide a nd conquer reduction strategy that splits multi-classes into binary grou ps and solves each individual problem recursively hence generating an un derlying decision list structure. The method is applied to assigning pr otein domains to folds. Experimental evaluation of the method demonstrat es the efficacy of the proposed approach to solving complex multi-class classification problems in bioinformatics. SUMMARY:W01> Multi-Class Protein Fold Recognition using Large Margin Logi c based Divide and Conquer Learning LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5232 DESCRIPTION:In protein sequence alignment algorithms\, a substitution mat rix of 20x20 alignment parameters is used to describe the rates of amino acid substitutions over time. Development and evaluation of most substi tution matrices including the BLOSUM family [1] was based almost entirel y on fully structured proteins. Structurally disordered proteins (i.e. p roteins that lack structure\, either in part or as a whole) that have be en shown to be very common in nature [2] have a significantly different amino acid composition than ordered (i.e. structured) proteins [3]. Furt hermore\, the sequence evolution rate is higher in unstructured as compa red to structured regions of proteins containing both structured and uns tructured regions [4]. These results cast doubt on appropriateness of th e BLOSUM substitution matrices for alignment of structurally disordered proteins [5].To address this problem\, we take into the account the conc ept of structural disorder by extending the alphabet for sequence repres entation from 20 to 2x20=40 symbols\, 20 for amino acids in disordered r egions and 20 for amino acids in ordered regions. A 40x40 substitution m atrix is required for alignment of sequences represented in the extended alphabet. Such an expanded matrix contains 20x20 submatrices that corre spond to matching ordered-ordered\, ordered-disordered\, and disordered- disordered pairs of residues. In this paper we describe an iterative pro cedure that we used to estimate such a 40x40 substitution matrix. The it erative procedure converged with stable results with respect to the choi ce of the sequences in the dataset. In the obtained 40x40 matrix we foun d substantial differences between the 20x20 submatrices corresponding to ordered-ordered\, ordered-disordered\, and disordered-disordered region matching. These differences provide evidence that for alignment of prot ein sequences that contain disordered segments\, the discovered substitu tion matrix is more appropriate than the BLOSUM substitution matrices. A t the same time\, the new substitution matrix is applicable for sequence alignment of fully ordered proteins as its order-order submatrix is ver y similar to a BLOSUM matrix. SUMMARY:W01> Protein Sequence Alignment and Structural Disorder: A Substi tution Matrix for an Extended Alphabet LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5233 DESCRIPTION:Experimental results often present a substantial fraction of missing and censored values. Here we propose a strategy to perform prin cipal component analysis under this specific incomplete information hypo thesis. Finally we show how to reconstruct the missing information in a way consistent with the experimental observations. SUMMARY:W01> Handling missing values and censored data in PCA of pharmaco logical matrices LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5234 DESCRIPTION:Recently\, the principles of graph theory are being adopted t o address molecular and chemical structures investigations such as 3D pr otein structure prediction and spatial motifs discovery. Proteins have b een parsed into graphs according to several approaches and methods and t hen studied based on graph theory concepts and data mining tools. In thi s paper we make a brief survey on the most used graph-based representati ons and we propose a naÔve method to help with the protein graph making since a key step of a valuable protein structure mining process is to b uild concise and correct graphs holding reliable information. We\, also\ , show that some existing and widespread methods present remarkable weak nesses and donít really reflect the real protein conformation. SUMMARY:W01> Comparing Graph-based Representations of Protein for Mining Purposes LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5235 DESCRIPTION:This paper gives a brief introduction into the biology of Tra nscription Factor Bindings Sites (TFBS's) and their importance. It discu sses the present methods employed in detecting them\, which are typicall y based on single-site frequencies\, their shortcomings and how new data is likely to revolutionise the area in the near future. The challenges associated with motif detection for methods that do not employ the stan dard approaches are addressed. Finally\, there is a walk through of so me of the available data sets for those interested in working on this pr oblem. SUMMARY:W01> Can we improve on the identification of Transcription Factor Binding Sites? LOCATION:Saint Michel END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T153000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T152000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5236 DESCRIPTION:Uncertainty is often inherent to data and still there are jus t a few data mining algorithms that handle it. In this paper we focus on how to account for uncertainty in classification algorithms\, in partic ular when data attributes should not be considered completely truthful f or classifying a given sample. Our starting point is that each piece of data comes from a potentially different context and\, by estimating cont ext probabilities of an unknown sample\, we may derive a weight that qua ntifies their influence. We propose a lazy classification strategy that incorporates the uncertainty into both the training and usage of classi fiers. We also propose uK-NN\, an extension of the traditional K-NN that implements our approach. Finally\, we illustrate uK-NN\, which is curre ntly being evaluated experimentally\, using a document classification to y example. SUMMARY:W11> Exploiting Contexts to Deal with Uncertainty in Classificati on LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T152000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T151000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5237 DESCRIPTION:The maximal clique enumeration (MCE) problem can be used to f ind very tightly-coupled collections of objects inside a network or grap h of relationships. However\, when such networks are based on noisy or u ncertain data\, the solutions to the MCE problem for several closely rel ated graphs may be necessary to accurately define the collections. Thus \, we propose an algorithm that e ciently solves the MCE problem on alte red\, or perturbed\, graphs. The algorithm utilizes the enumeration of a baseline graph and identifies only those maximal cliques that the pertu rbation adds and/or removes. We detail the algorithm and the underlying theory required to guarantee correctness. Further\, we report average ru ntime speedups of 7 and 9 for our algorithm over traditional enumeration techniques in the cases of adding and removing edges\, respectively\, f rom graphs constructed from protein interaction data. SUMMARY:W11> On Perturbation Theory and an Algorithm for Maximal Clique E numeration in Uncertain and Noisy Graphs LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T150500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T145000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5238 DESCRIPTION:Mining of frequent patterns is one of the popular knowledge d iscovery and data mining (KDD) tasks. It also plays an essential role in the mining of many other patterns such as correlation\, sequences\, and association rules. Hence\, it has been the subject of numerous studies since its introduction. Most of these studies find all the frequent patt erns from collection of precise data\, in which the items within each da tum or transaction are definitely known and precise. However\, there are many real-life situations in which the user is interested in only some tiny portions of these frequent patterns. Finding all frequent patterns would then be redundant and waste lots of computation. This calls for co nstrained mining\, which aims to find only those frequent patterns that are interesting to the user. Moreover\, there are also many real-life si tuations in which the data are uncertain. This calls for uncertain data mining. In this paper\, we propose an algorithm to efficiently find cons trained frequent patterns from collections of uncertain data. SUMMARY:W11> Efficient Algorithms for Mining Constrained Frequent Pattern s from Uncertain Data LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T144000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5239 DESCRIPTION: SUMMARY:W11> Invited Talk "Managing and Mining Uncertain Data: What Might We Do Better?" LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T145500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T144000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5240 DESCRIPTION: SUMMARY:W11> Efficient Algorithms for Mining Constrained Frequent Pattern s from Uncertain Data LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T151000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T145500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5241 DESCRIPTION: SUMMARY:W11> Identifying Graphs from Noisy and Incomplete Data LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T163500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T160000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5242 DESCRIPTION: SUMMARY:W11> Invited talk "Querying and Mining Uncertain Data: Methods\, Applications\, and Challenges" LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T165000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T163500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5243 DESCRIPTION: SUMMARY:W11> Learning from Data with Uncertain Labels by Boosting Credal Classifiers LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T173000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T140000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5244 DESCRIPTION: SUMMARY:W11> Using Uncertain Chemical and Thermal Data to Predict Product Quality in a Casting Process LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T171500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T170500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5245 DESCRIPTION: SUMMARY:W11> Lazy Naive Credal Classifier LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T172500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T171500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5246 DESCRIPTION: SUMMARY:W11> Decision Support and Profit Prediction for Online Auction Se llers LOCATION:Les Invalides B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T094500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T090500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5247 DESCRIPTION: SUMMARY:W08> Invited Talk "Tensor Decompositions and Applications: a Surv ey" LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T100000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T094500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5248 DESCRIPTION: SUMMARY:W08> Multi-Way Set Enumeration in Real-Valued Tensors LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T102000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5249 DESCRIPTION: SUMMARY:W08> Invited Talk "Factorizing Matrices with Missing Entries: Alt ernative Approaches" LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T111500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T110000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5250 DESCRIPTION: SUMMARY:W08> A Spectral-based Clustering Algorithm for Categorical Data U sing Data Summaries LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T113000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T111500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5251 DESCRIPTION: SUMMARY:W08> Sequential Latent Semantic Indexing LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T120000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T113000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5252 DESCRIPTION: SUMMARY:W08> Invited Talk: "Tensors and n-d Arrays: Mathematics of Arrays \, Psi-Calculus\, and Composition of Tensor and Array Operations" LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T121000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T120000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5253 DESCRIPTION: SUMMARY:W08> Recent Advances in Tensor Decomposition LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T122500 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T121000 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5254 DESCRIPTION: SUMMARY:W08> Accuracy of Distance Metric Learning Algorithms LOCATION:St Germain des Prés B END:VEVENT BEGIN:VEVENT DTEND;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T124000 DTSTART;TZID=Europe/Paris;VALUE=DATE-TIME:20090628T122500 DTSTAMP;VALUE=DATE-TIME:20120516T153740Z UID:5255 DESCRIPTION: SUMMARY:W08> Efficient Computation of PCA with SVD in SQL LOCATION:St Germain des Prés B END:VEVENT END:VCALENDAR