Bayesian collaborative competitive filtering

Abstract

A recommendation system is inherently an interactive system that involves recommending action of the system and according responses from the user. This interaction process had not been fully exploited for learning recommender models until the choice-based recommendation system was proposed. Recently, a choice-based recommender algorithm called Collaborative Competitive Filtering (CCF) was proposed. CCF exploits the information generated while the system and user interact with each other. CCF uses a multinomial logit model to describe user`s choice behavior occurring in the user-system interactions. CCF showed superb predictive capability in several large-scale real-world datasets, outperforming other rating-based recommender algorithms. The main drawback of CCF, however, is the requirements of manual complexity control. In CCF, regularization parameters have to be manually tuned to make the model generalize well on unseen data. The computational costs of the procedure, however, are often prohibitive, since it requires training a multitude of models.

In this paper, we introduce Bayesian Collaborative Competitive Filtering (BCCF) which is a fully Bayesian treatment of the CCF model. Unlike CCF, BCCF automatically controls model capacity by integrating over all model parameters and hyperparameters. We also introduce an MCMC-based inference algorithm for the BCCF model. Experiments on large-scale real-world datasets demonstrate that BCCF outperforms CCF and other rating-based recommender algorithms.