A1731
Title: Scalable data augmentation for high-dimensional Bayesian regression and classification
Authors: Gyuhyeong Goh - Kyungpook National University (Korea, South) [presenting]
Abstract: With the increasing prevalence of high-dimensional data, shrinkage priors have become standard tools in Bayesian regression modeling. As most shrinkage priors used for sparse regression belong to the class of normal-scale mixtures, Gibbs sampling has attracted considerable attention for performing exact Bayesian inference. However, when the dimension of the coefficient vector is large, the conventional Gibbs sampler under normal-scale mixture priors becomes computationally expensive. A scalable data augmentation approach enables exact Bayesian inference for high-dimensional regression and classification under a general class of normal-scale mixture priors, including spike-and-slab and horseshoe priors. A distinguishing feature of the proposed posterior sampler is that its per-iteration computational complexity can be made to scale linearly with the size of the design matrix.
