Towards Context-Aware Social Recommendation via Trust Networks

Utilizing social network information to improve recommendation quality has recently attracted much attention. However, most existing social recommendation models cannot well handle the heterogeneity and diversity of the social relationships (e.g., different friends may have different recommendations on the same items in different situations). Furthermore, few models take into account (non-social) contextual information, which has been proved to be another valuable information source for accurate recommendation. In this paper, we propose to construct trust networks on top of a social network to measure the quality of a friend's recommendations in different contexts. We employ random walk to collect the most relevant ratings based on the multi-dimensional trustworthiness of users in the trust network. Factorization machines model is then applied on the collected ratings to predict missing ratings considering various contexts. Evaluation based on a real dataset demonstrates that our approach improves the accuracy of the state-of-the-art social, context-aware and trust-aware recommendation models by at least 5.54% and 9.15% in terms of MAE and RMSE respectively.