A Bayesian Exponential Family Harmonium (BEFH) model is presented for topical modeling of text and multimedia data, and for "posterior" latent semantic projection of such data for subsequent data mining tasks. BEFHs are a Bayesian approach to inference and learning with the recently proposed EFH models and their variants, which enables smoothed, robust estimation of the topic-attribute coupling coefficients that are reminiscent of the smoothed topical word-probabilities in the latent Dirichlet Allocation (LDA) model. The Langevin algorithm conjoint with an MCMC scheme is applied for posterior inference with BEFH. An empirical Bayes method is also developed to estimate the hyperparameters.

18 pages

This page maintained by reports@cs.cmu.edu