A1394
Title: Generalized reparametrized variational Bayes with skew-symmetric normalization
Authors: Aoxiang Chen - National University of Singapore (Singapore) [presenting]
Siew Li Linda Tan - National University of Singapore (Singapore)
Abstract: Bayesian hierarchical models with high-dimensional latent structure require scalable posterior approximations that retain dependence and non-Gaussian features. Mean-field variational inference (MFVI) is fast, but can be inaccurate when local variables are correlated, skewed, or strongly coupled with global variables. KNorm-RVB, a generalized reparametrized variational Bayes framework for latent Gaussian and latent non-Gaussian models with sparse local precision, normalizes the local latent vector and reduces skewness, moving its posterior closer to standard Gaussian. A novel $K$-component skew-symmetric construction based on blockwise reflection groups enforces zero local mean and zero local-global covariance, making MFVI highly effective on the reparametrized model. A zero-mean Gaussian variational family is used for the reparametrized local variables and flexible closed skew-normal families for the remaining variables. Across generalized linear mixed models, spatial simultaneous autoregressive models, and stochastic volatility models, KNorm-RVB improves both marginal and multivariate accuracy over existing methods.
