Conjugate priors for Gaussian emission plsa recommender systems

Collaborative filtering for recommender systems seeks to learn and predict user preferences for a collection of items by identifying similarities between users on the basis of their past interest or interaction with the items in question. In this work, we present a conjugate prior regularized extension of Hofmann's Gaussian emission probabilistic latent semantic analysis model, able to overcome the over-fitting problem restricting the performance of the earlier formulation. Furthermore, in experiments using the EachMovie and MovieLens data sets, it is shown that the proposed regularized model achieves significantly improved prediction accuracy of user preferences as compared to the latent semantic analysis model without priors.