Jiawei (Peter) Liu

Title: Training LDA models with Measure Preserving and Discrete Maps
Date: Monday, August 17th
Time: 3:00pm
Location: Zoom
Supervised by: Owen Ward & Liangliang Wang

Abstract: Latent Dirichlet Allocation (LDA) is a Bayesian topic model commonly fitted by collapsed Gibbs sampling. This thesis investigates replacing conventional categorical resampling with the Measure-Preserving and Discrete (MAD) map. MAD-based algorithms are developed for both LDA and LDA with covariates (LDAcov), with slice sampling used for the continuous LDAcov model parameters. Experiments on Associated Press articles, simulated LDAcov data, and NEON soil-bacteria data show that the MAD-based LDA sampler achieved lower held-out perplexity in the reported experiment, while MAD- and inversion-based LDAcov samplers produced similar parameter recovery and computational cost. These results support MAD updates as a practical alternative for topic-model inference.