Dynamic Compound-Poisson Factorization

Abstract

Collaborative Filtering (CF) analyzes previous user-item interactions in order to infer the latent factors that represent user preferences and item characteristics. However, most current collaborative filtering algorithms assume that these latent factors are static while user preferences and item perceptions drift over time. In this paper, we propose a novel Bayesian Dynamic Matrix Factorization model based on Compound Poisson Factorization that models the smoothly drifting latent factors as Gamma chains. We provide a novel approach to Gamma chains to guarantee their conjugacy and numerical stability. We then provide a scalable inference algorithm to learn the parameters. We finally apply our model to timestamped ratings datasets such as Netflix, Yelp, LastFm where we achieve higher predictive accuracy than state-of-the-art static factorization models.