Variable Selection for Bilinear Regression with Matrix Covariates

• Activity Number: 205 - New Direction for Model Selection in Big Data
• Type: Invited
• Date/Time: Monday, July 31, 2017 : 2:00 PM to 3:50 PM
• Sponsor: IMS
• Author(s): Dan Yang*
• Companies: Rutgers University

Abstract

Rapid development of modern technology has generated numerous high-dimensional datasets with matrix-valued covariates. Application of traditional LASSO after vectorization of these matrix-valued covariates leads to inferior estimation and selection results. An existing proposal that preserves the matrix structure through bilinear regression adopts a penalty that essentially assumes the same degree of sparsity along the rows and columns, which is often violated by real data. We introduce a novel penalty that solves this problem. The resulting method is minimax rate optimal and outperforms the existing one in the simulation studies.