Talk by: Arindam Banerjee
Yahoo! Academic Relations and The Columbia University Center for Computational Learning Systems Present:
Title: Probabilistic Models for Analyzing Large Sparse Matrices
Who: Arindam Banerjee
When: Friday December 3rd, 2010
Where: Interschool Lab, 7th Floor CEPSR
Data matrices arising out of diverse domains ranging from collaborative filtering to internet applications and forest ecology have characteristics which severely challenge traditional approaches to matrix analysis. Such matrices represent large scale data in high dimensions and most entries in such matrices are missing. In this talk, I will discuss two families of probabilistic graphical models designed for effective analysis and missing value prediction for such large sparse matrices. The first family relies on the co-clustering structure in a matrix. Instead of finding one co-clustering, the model maintains a distribution over all co-clusterings. The second family relies on the factorization structure in a matrix. Instead of trying to obtain one good factorization, the model maintains a probability distribution over all factorizations. Both models can naturally handle sparse matrices since only observed entries are assumed to be generated from the model. Further, the models are modular so that side information about rows and columns can be easily incorporated into the analysis. I will discuss efficient approximate inference algorithms for learning and missing value prediction in both models. I will present several experimental results to illustrate the effectiveness and capabilities of these models, and their competitive advantage over existing methods for matrix analysis. I will also discuss how related probabilistic graphical models can be developed for a wide range of problems in classification and clustering with state of the art performance.
Arindam Banerjee is an Assistant Professor and a McKnight Land Grant Professor in the Department of Computer Science & Engineering, and a Resident Fellow in the Institute on the Environment (IonE) at the University of Minnesota, Twin Cities. He received his Ph.D. from the University of Texas at Austin in 2005, where his dissertation was nominated for the best dissertation award. His research interests are in Machine Learning, Data Mining, Information Theory, Convex Analysis and Optimization, and their applications in complex real world problems including those in Text and Web Mining, Social Network Analysis, Healthcare, Bioinformatics, Climate and Environmental Sciences, and Finance.
He has won several awards including the NSF CAREER award in 2010, the McKnight Land-Grant Professorship at the University of Minnesota, Twin Cities, 2009-2011, the J. T. Oden Faculty Research Fellowship from the Institute for Computational Engineering and Sciences (ICES), University of Texas at Austin, 2006, and the prestigious IBM PhD fellowship for the academic years 2003-2004 and 2004-2005. He has also won several awards for his publications, including the Best Paper Award at the SIAM International Conference on Data Mining (SDM), 2004, the Best Research Paper Award under University Cooperative Society Research Excellence Awards, University of Texas at Austin, 2005, and the Best of SIAM Data Mining (SDM) Award at the SIAM International Conference on Data Mining, 2007