Nonparametric Bayes Multi-Task Multi-View Learning

• Activity Number: 92
• Type: Invited
• Date/Time: Sunday, August 4, 2013 : 8:30 PM to 10:30 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Angela Schoergendorfer*+ and Hongxia Yang
• Companies: IBM T.J. Watson Research Center and IBM T.J. Watson Research Center
• Keywords: matrix dirichlet process ; machine learning ; big data

Abstract

Large-scale machine learning problems often consist of multiple learning tasks and heterogeneous feature sets ('views') that may be shared across tasks. Examples for such multi-task multi-view problems are cross-lingual document classification, cross-domain sentiment analysis, or web image classification. Traditional approaches have addressed multi-task learning separately from multi-view learning. More recently, joint models of the two types of heterogeneity have been proposed to improve learning performance, particularly with limited training data. We present a nonparametric Bayes framework for learning relationships among tasks and views simultaneously. Task relatedness is modeled via a Gaussian process, while view relatedness is modeled by a matrix Dirichlet process. We expand previous formulations of the model by allowing for views to retain different dimensions. We propose an efficient Gibbs algorithm and present experimental results comparing performance of our proposed method to other methods that have been proposed for the multi-task multi-view problem.