NOTAM 2 : Nonparametric Bayes Multi-Task Multi-View Learning
Hongxia Yang, Jingrui He · 2013
Heterogeneous learning refers to addressing problems with multiple types of heterogeneity, e.g., task heterogeneity, view heterogeneity, etc. It finds abundant applications in cross-lingual document classification, cross-domain sentiment analysis, web image classification, etc. Traditional approaches handle different types of heterogeneity separately via multi-task learning, multi-view learning, etc. More recently, researchers start to jointly model different types of heterogeneity in order to improve the learning performance with limited training data. In this paper, we advance stateof-the-art in heterogeneous learning by jointly modeling task and view relatedness via nonparametric Bayes method. To be specific, we model task relatedness using normal penalty with sparse covariances to couple multiple tasks and view relatedness using matrix Dirichlet process. We also propose NOTAM 2 algorithm, which is based on an efficient Gibbs algorithm. Experimental results demonstrate the effectiveness of NOTAM 2 .