Neural Information Processing - Letters and Reviews
Vol. 12, Nos. 1-3, January-March 2008
α-Bayesian Collaboration of Multiple Predictors and its Applications to Hybrid Recommendation and User Modeling
Jun-ichiro Hirayama, Takashi Takenouchi
Graduate School of Information Science, Nara Institute of Science and Technology,
Takayama 8916-5, Ikoma, Nara
Ricoh Company, Ltd.
Graduate School of Informatics, Kyoto University
In recent widespread areas of machine learning or data mining applications, one often should deal with multiple statistical prediction tasks simultaneously, each of which cannot be successful by itself due to limitation of data amount. Among several approaches to utilizing task relationship, a naive but still important approach is to separately train taskspecific predictors in advance and then integrate them at the time of prediction. In this study, we propose a general framework of realizing such a “collaborative prediction” mechanism, specifically based on an existing generalization of Bayesian predictive distribution using the α-divergence. We also propose a novel hybrid method of collaborative and content-based recommendations, under the proposed framework. We demonstrate the effectiveness of the proposed method by using two kinds of real datasets; one is the Movielens benchmark collection related to recommendation, and the other is a real log dataset of shared electronic devices related to a particular problem of user modeling.
Keywords – Bayesian predictive distribution, Collaborative prediction, Movielens benchmark