Online Bayesian Analysis and Experimental Design for Massive Datasets

Electrical and Computer Engineering

Speaker: Lawrence Carin
Monday, February 13, 2012
4:00 PM to 5:00 PM


1049 Duncan Hall
Rice University
6100 Main St
Houston,Texas,USA


Abstract: With the development of the internet, and networked systems, there is a deluge of data available for analysis. Application areas include analysis of social networks and recommender systems. In this talk we address this challenge from a Bayesian perspective, developing techniques for analysis of static and dynamic data of very high dimension. In this framework we learn a low-dimensional latent model; this low-dimensional framework is essential for handling data with a high degree of missingness, allowing accurate inference of missing data. Additionally, we consider information-theoretic methods that quantify which missing data would be most informative to acquire, in the context of inferring the low-dimensional model. Ideas from submodularity are employed to allow acquisition of multiple missing samples at a time. In addition, we develop a fast online variational Bayesian tool for performing computations, allowing online modeling of data of massive dimension. Example results for static and dynamic data are presented for several real-world applications.