Bayesian Multinomial Logistic Regression for Numerous Categories

While multinomial logistic regression is a useful tool for classification among multiple categories, the posterior sampling of Bayesian implementations is computationally burdensome when the number of categories is large. In this paper, we show that the appropriate data augmentation technique provides faster posterior sampling than alternatives in the literature. This speed up comes from two sources: simpler posterior conditional distributions on the coefficients and the ability to parallelize parameter draws. In simulation studies, we demonstrate that the effective sampling rate of our posterior sampling approach is double that of competing methods when working with a large number of categories, even without parallelized computations. Furthermore, this computation time only increases linearly as the number of categories increases. Our corresponding R package is available on Github.
Comment: 10 pages, 2 figures. R package available at https://github.com/kylemcevoy/BayesMultiLogit