Talk Title
Simultaneous Predictor and Response Dimension Reduction for Multivariate Categorical Regression
Abstract
In multivariate categorical regression, the number of parameters can increase rapidly with the number of predictors, response variables, and categories within each response. This poses challenges in estimation, interpretation, and computational efficiency. We propose to employ tensor decompositions, specifically Tucker and Tensor Train decompositions, to exploit low-rank structures in both predictor and responses. Additionally, we explore the impact of different tensor decompositions on the interpretability of the fitted models, with a particular focus on settings where the responses are highly dependent. We develop algorithms to estimate the models under these decompositions and conduct simulation studies to evaluate their performance.
Bio
Yisen Jin is a Ph.D. candidate in Statistics at the University of Florida, working under the supervision of Dr. Aaron Molstad. His research interests lie in multivariate analysis, dimension reduction, and numerical optimization.