A1205
Title: Joint low-rank and individual sparse modeling for multi-response matrix-variate trace regression
Authors: Daeyoung Ham - The University of Texas at San Antonio (United States) [presenting]
Brad Price - West Virginia University (United States)
Abstract: Multi-response regression with matrix-valued predictors is studied. A unified joint-structure learning approach is proposed that separates signal shared across responses from response-specific heterogeneity, enabling simultaneous borrowing of strength and individualized adaptation in high dimensions. A convex regularized estimator is introduced that combines low-rank structure for the common component with sparse structure for the individual deviations, and a scalable optimization algorithm based on alternating direction methods is developed. The same framework extends naturally from continuous outcomes to multivariate generalized linear models such as multivariate logistic regression while retaining the same decomposition and computational strategy.
