Abstract
Reproducing kernel Hilbert space(RKHS) is a popular tool used in high dimensional data analysis. For example, smoothing spline and SVM (support vector machine) are special cases of RKHS. In this talk, I'll present a fully Bayesian framework and theory that coherently embed kernel regression/classification in a general nonparametric model. Key practical features of our approach include the use of shrinkage priors to address problems of ``large p'', the use of mixture priors for feature selection, coherent updating as sample sizes change, and an understanding of so-called ``unlabelled'' data.

Bio:
Rosie Jones is a Senior Research Scientist at Yahoo! where her recent work includes research on query log analysis. Her research interests include information retrieval, machine learning, statistical natural language processing and time series analysis. She received her PhD from the School of Computer Science at Carnegie Mellon University in 2005 under the supervision of Tom Mitchell, where her doctoral thesis was titled Learning to Extract Entities from Labeled and Unlabeled Text. In 2005 she co-organized the SIGIR workshop on lexical cohesion and information retrieval, and in 2003 she co-organized the ICML workshop on The Continuum from Labeled to Unlabeled Data in Machine Learning and Data Mining.

Bio:
Xuehua Shen is a Ph.D. candidate in the Computer Science Department of University of Illinois at Urbana-Champaign. His main research interests are in information retrieval, database, and data mining. In addition, he is also broadly interested in some other areas in computer science, especially machine learning, human computer interaction, and information privacy. He has published papers in conferences of ACM SIGIR, ACM CIKM, ACM SIGKDD, WWW, and IEEE ICDE. He received his B.S. in Computer Science from Nanjing University, China in 1999. He did summer interns at Microsoft Research Redmond and IBM T. J. Watson Research Center. Recently, he served on the program committee of Symposium on Information Interaction in Context (IIiX'2006). He has two patents related with personalized search filed or in preparation.

Bayesian Logistic Regression for Text Classification
Spearker: David Lewis
Abstract:
Bayesian logistic regression allows incorporating task knowledge through model structure and priors on parameters. I will discuss content-based text categorization and authorship attribution using 1) priors that control sparsity and sign of parameters, 2) priors that incorporate domain knowledge from reference books and other texts, and 3) the use of polytomous (1-of-k) dependent variables. All experiments were performed with our open source programs, BBR and BMR, which can fit models with millions of parameters. (Joint work with David Madigan, Alex Genkin, Aynur Dayanik, Dmitriy Fradkin, and Vladimir Menkov at Rutgers University and DIMACS.)

Bio:
Dave Lewis is an entrepreneur and consulting computer scientist based in Chicago, IL. He consults at, and is building another company at, the intersection of information retrieval, data mining, and natural language processing. He previously held research positions at AT&T Labs, Bell Labs, and the University of Chicago, and received his Ph.D. in Computer Science from the University of Massachusetts. He was recently elected a Fellow of the American Association for the Advancement of Science.

A General Model for Retrieval and Filtering
Speaker: Javed Mostafa (from Indiana Univ.)
Abstract:
Information retrieval and filtering (IR/IF) over dynamic environments such as the WWW poses many challenges. At the root of the challenges lie the twin factors of data diversity and data evolution. These two factors are especially problematic for content representation and classification. With a growing number of people relying on web search applications to support everyday tasks, another emerging challenge is providing highly focused results (i.e., personalization) while accommodating diverse user-interests and interest drifts. In fact, the key challenges associated with representation, classification, and personalization are interdependent and the likelihood of achieving progress in any one of these areas can be increased by focusing on them together -- as part of a single information services framework.

In this talk, I will present and discuss such a framework, known as the multilevel information services model. I will demonstrate how the modular nature of the framework permits the isolation and study of key functions that are more basic than IR/IF and establish their individual and combined impact on the effectiveness and efficiency of information services. I will also present results from a series of experiments examining how factors such as the representation method (e.g., automated feature generation vs. manual schemes), classifiers (e.g., supervised vs. unsupervised), nature and size of content streams, level of user interaction, and rate of interest change influence system performance.

Bio:
Dr. Javed Mostafa is the Victor H. Yngve Associate Professor of Information Science and Associate Professor of Informatics at Indiana University, Bloomington. He also has formal faculty affiliations with the Cognitive Science Program in Bloomington (core faculty) and the Computer & Information Science Department (adjunct) in Indianapolis. He is widely published in the information retrieval area. His current research focuses on secure and personalized access to health information. Dr. Mostafa is an associate editor of the ACM Transactions on Information Systems.

Bio
Sanjay Chawla is an Associate Professor in the School of Information Technologies, University of Sydney, Australia. He currently serves as an Associate Head of the School. Sanjay's research work is in the area of data mining with an emphasis on spatial data and more recently on outlier detection. Sanjay is a co-author on the text "Spatial Databases: A Tour." Before joining the University of Sydney, Sanjay worked for Vignette Corporation in Boston (2000- 2002); was a postdoc in the Department of Computer Science, University of Minnesota (1997- 2000) and an industrial postdoc at the Institute for Mathematics and its Application (IMA), University of Minnesota(1995-1997). He received his PhD from the University of Tennessee, Knoxville in 1995. He is a member of ACM, SIAM and AMS.

Information is the key to effective knowledge and decision-making. However, in order to be useful, information must be accessible. Information graphics (non-pictorial graphics such as bar charts, line graphs, and pie charts) pose accessibility problems:
1) although much attention has been devoted to the summarization, categorization, and retrieval of text documents and images, almost no attention has been given to information graphics,
2) individuals with impaired eyesight lack effective methods for accessing the information provided by information graphics.

The overwhelming majority of information graphics that appear in newspapers, magazines, and formal reports have a communicative goal or message that they are intended to convey. This talk will first present our corpus study showing that the message conveyed by an information graphic in a multimodal document is generally not contained in the article's text. The talk will then present our research on identifying the primary intended message of an information graphic; this message can then be used as the core of a summary of the graphic for digital libraries or as part of an interactive natural language system for blind individuals.

Our approach is to view information graphics as a form of language containing communicative signals. We will discuss the kinds of communicative signals appearing in information graphics, and describe how we exploit these in a Bayesian network for recognizing the graphic's intended message. We will present our implemented system along with the results of evaluation experiments that demonstrate the effectiveness of our methodology.

This research has been done jointly with Stephanie Elzer (former graduate student at the University of Delaware and currently an assistant professor at Millersville University). Other contributors to the work include Dan Chester, Seniz Demir, and Peng Wu.

Short Bio:
Sandra Carberry is a Professor and former Chair of the Department of Computer Science at the University of Delaware. She received a BS degree from Cornell University, an MS degree from Rice University, and a PhD from the University of Delaware, and she was a Member of Technical Staff at Bell Laboratories. She served as Secretary and Member of the Executive Board of the Association for Computational Linguistics from 2001-2006. She is currently on the Editorial Board of the "Journal of Dialog Systems" and "User Modeling and User-Adapted Interaction". Her research focuses on issues of effective communication.

Bayesian Logistic Regression for Text Classification
Spearker: David Lewis
Abstract:
Bayesian logistic regression allows incorporating task knowledge through model structure and priors on parameters. I will discuss content-based text categorization and authorship attribution using 1) priors that control sparsity and sign of parameters, 2) priors that incorporate domain knowledge from reference books and other texts, and 3) the use of polytomous (1-of-k) dependent variables. All experiments were performed with our open source programs, BBR and BMR, which can fit models with millions of parameters. (Joint work with David Madigan, Alex Genkin, Aynur Dayanik, Dmitriy Fradkin, and Vladimir Menkov at Rutgers University and DIMACS.)

Bio:
Dave Lewis is an entrepreneur and consulting computer scientist based in Chicago, IL. He consults at, and is building another company at, the intersection of information retrieval, data mining, and natural language processing. He previously held research positions at AT&T Labs, Bell Labs, and the University of Chicago, and received his Ph.D. in Computer Science from the University of Massachusetts. He was recently elected a Fellow of the American Association for the Advancement of Science.