A Regularized Recommendation Algorithm with Probabilistic Sentiment-Ratings

Recommender systems (RS) provide personalized suggestions based on users' past behavior and/or similarities between users' and products' profiles. Although we observed a high interest in the research community over RS algorithms these commonly overlook users' opinions. In this paper, we research the inclusion of sentiment knowledge in RS to improve the overall quality of recommendations. In contrast to similar approaches, we propose a matrix factorization with a new factor to regularize probabilistic ratings. A sentiment analysis algorithm implementing a multiple Bernoulli classification computes these probabilistic ratings. The combination of a regularization factor with probabilistic ratings offers a general framework capable of embedding multiple sources into a theoretical well-founded matrix factorization algorithm. Experiments show that with an evaluation on a dataset with 1.7 million reviews we have successfully introduced a novel approach to incorporate on a RS with inferred rating based in a sentiment analysis framework. Also, replacing explicit ratings by probabilistic inferred ratings the RS performance improves, thus, our proposed framework is able to better accommodate the uncertainty of users explicit rating.