Bayesian probabilistic tensor factorization for recommendation and rating aggregation with multicriteria evaluation data

• Published in: Expert systems with applications 2019-10, Vol.131, p.1-8
• Main Authors: Morise, Hiroki, Oyama, Satoshi, Kurihara, Masahito
• Format: Article
• Language: English
• Subjects: Agglomeration; Approximation; Bayesian analysis; Bayesian probabilistic models; Collaborative filtering; Computer simulation; Markov analysis; Markov chains; Monte Carlo simulation; Multi-criteria rating; Multiple criterion; Probabilistic models; Rating aggregation; Ratings; Recommendation; Recommender systems; Tensors
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2019.04.044

•To the best of our knowledge, our work is the first attempt to apply Bayesian probabilistic tensor factorization to multicriteria recommendation. Our model, which we call “Bayesian probabilistic tensor factorization for multicriteria (BPTF-MC),” predicts the overall rating and the rating from each viewpoint simultaneously. It does this by using multicriteria latent features as additional factors.•The BPTF-MC model enables the prediction of ratings for items by each user and of aggregated ratings from the evaluations of a small number of users.•Experimental results for the Rakuten public datasets show that the BPTF-MC model achieves better performance than single-criterion models and low-rank tensor factorization models for both recommendation and rating aggregation. Ratings by users on various items such as products and services have become easily available on the Web. Also available in many cases, in addition to an overall rating for each item by each user, are multicriteria ratings from different viewpoints. Our previous study showed that multicriteria rating approaches performed better than single-criterion ones for both recommendation and rating aggregation. We have now formulated a Bayesian probabilistic model for multicriteria evaluation as an alternative to low-rank approximation. We evaluated the performance of this model, in which model capacity is controlled by integrating over all model parameters, and investigated whether it can be made to work more efficiently by using a Markov chain Monte Carlo method for both recommendation and rating aggregation. It performed better than low-rank approximation methods that obtain a maximum a posteriori estimate by fitting to the data.