A0878
Title: Laplace variational inference for Bayesian envelope models
Authors: Seunghyeon Kim - Chonnam National University (Korea, South) [presenting]
Kwangmin Lee - Chonnam National University (Korea, South)
Yeonhee Park - Sungkyunkwan University (Korea, South)
Abstract: Envelope models achieve efficient estimation by employing dimension reduction in multivariate regression. However, traditional Markov chain Monte Carlo (MCMC)-based Bayesian methods are computationally slow. A Laplace variational inference (LVI) approach is proposed for Bayesian envelope models. The approach involves a novel reparameterization of the posterior distribution that enables direct computation of the second derivative required by LVI. Theoretical results are provided, showing that the variational distribution derived via LVI asymptotically converges to the distribution obtained by standard variational inference (VI). Extensive simulation studies and a real data analysis illustrate the superior computational efficiency, estimation accuracy, and model selection performance of the proposed method compared to existing approaches.
