Probabilistic Topic Model for Hybrid Recommender Systems: A Stochastic Variational Bayesian Approach

Abstract

Internet recommender systems are popular in contexts that include heterogeneous consumers and numerous products. In such contexts, product features that adequately describe all the products are often not readily available. Content-based systems therefore rely on user-generated content such as product reviews or textual product tags to make recommendations. In this paper we develop a novel covariate-guided heterogeneous supervised topic model that uses product covariates, user ratings and product tags to succinctly characterize products in terms of latent topics, and specifies consumer preferences via these topics. Recommendation contexts also generate big data problems stemming from data volume, variety and veracity, as in our setting that includes massive textual and numerical data. We therefore develop a novel stochastic variational Bayesian (SVB) framework to achieve fast, scalable and accurate estimation in such big data settings, and apply it on a MovieLens dataset of movie ratings and semantic tags. We show that our model yields interesting insights about movie preferences and predicts much better than a benchmark model that only uses product covariates. We showcase how our model can be used for targeting recommendations to particular users and illustrate its use in generating personalized search rankings of relevant products.

Keywords: Hybrid Recommendation Models, Personalized Search, User-Generated Content, Probabilistic Topic Models, Big Data, Scalable Inference, Stochastic Variational Bayes

JEL Classification: C01, C11, C13, C55, M31