Title: Bayesian sparse factor models with overlapping blocks
Authors: Ilsang Ohn - Seoul National University (Korea, South) [presenting]
Yongdai Kim - Seoul National University (Korea, South)
Abstract: Sparse factor models have proven useful for describing dependency in high-dimensional data. Whereas sparsity of the factor loading matrix improves both interpretability and predictive performance, it may yield a covariance matrix having too many zero correlations, which is not desirable in many situations including genetic data where the variables are expected to be highly correlated. The aim is to propose a Bayesian sparse factor model with the corresponding covariance having overlapping block structure. The proposed factor model is able to capture strong dependency between random variables with the relatively small number of parameters. We introduce a novel prior distribution on the factor loading matrix, which provides lots of flexibility and enables scalable posterior computation. We show on a number of datasets that our model outperforms other competitors.