Modelling socially-influenced conditional preferences over feature values in recommender systems based on factorised collaborative filtering

Abstract

Abstract Recommender systems have gained much attention due to their great commercial benefits in electronic markets. The quality of the recommendations depends on the quality of the preference model extracted by the recommender system. Recently, latent factor models based on probabilistic matrix factorisation have gained much attention, owing to their superior accuracy over traditional recommender systems and their great efficiency. Although latent factor models are very efficient, they mostly ignore the user preferences over different item feature values. For example, they assume that lower prices are preferred by all users. However, there may be users who believe that a high price comes with better quality or more prestige. Furthermore, according to homophily and social influence in social sciences, similar users in a social network tend to acquire similar tastes through social interactions. Therefore, all components of human preferences including feature value discrepancies are subject to social influence. Moreover, most of the latent factor models ignore the possible dependencies that naturally exist between item features. To tackle these problems, in this paper we propose two novel latent factor models incorporating socially-influenced feature value discrepancies, and socially-influenced conditional feature value discrepancies. We test the accuracy of the proposed methods on three well-known benchmark datasets. The extensive experiments show that the proposed method achieves significantly higher accuracies than all state of the art traditional, latent factor, and social recommendation models.