Recommender Systems with Social Regularization

“Traditional recommender systems always ignore social relationships among users. But in our real life, when we are asking our friends for recommendations of nice digital cameras or touching movies, we are actually requesting verbal social recommendations. Social recommendation is a daily occurrence, and we always turn to our friends for recommendations. Hence, in order to improve recommender systems and to provide more personalized recommendation results, we need to incorporate social network information among users.

Recently, based on the intuition that users’ trust relations can be employed to enhance traditional recommender systems, a few trust-aware recommendation methods have been proposed [2, 22, 23, 25, 27]. These methods utilize the inferred implicit or observed explicit trust information to further improve traditional recommender systems. Trust-aware recommender systems move an important step forward in the research of recommender systems. However, to achieve the goal of “social recommendation”, these approaches still have several inherent limitations and weaknesses that need to be addressed.

First of all, “trust relationships” are different from “social friendships” in many aspects. Typically, on theWeb, when a user ui likes a review issued by another user ut, user ui probably will add user ut to his/her trust list. This process of trust generation is a unilateral action that does not require user ut to confirm the relationship. This also indicates that user ui does not need to even know user ut in the real life. “Social friendships” refer to the cooperative and mutual relationships that surround us, such as classmates, colleagues, or relatives, etc. Lots of social networking Web sites, like Facebook and Orkut, are designed for online users to interact and connect with their friends in the real life. From the definition, we can see that trust-aware recommender systems cannot represent the concept of “social recommendation”, since the idea of “social recommendation” anticipates to improve recommender systems by incorporating a social friend network.

Secondly, trust-aware recommender systems are based on the assumption that users have similar tastes with other users they trust. This hypothesis may not always be true in social recommender systems since the tastes of one user’s friends may vary significantly. Some friends may share similar favors with this user while other friends may have totally different tastes. Hence, trust-aware recommendation algorithms cannot be directly applied to generate recommendations in social recommender systems.

Thirdly, due to the rapid growth of Web 2.0 applications, online users spend more and more time on social network related applications since interacting with real friends is the most attractive activity on the Web. On the contrary, only few online systems, like Epinions, have implementations of trust mechanism. Thus, in order to provide more proactive and personalized recommendation results to online users, we should pay more attention to the research of social recommendation, in addition to the existing research of trustaware recommendation.

In this paper, aiming at solving the problems mentioned above, we propose two social recommendation methods that utilize social information to improve the prediction accuracy of traditional recommender systems. More specifically, the social network information is employed in designing two social regularization terms to constrain the matrix factorization objective function. Moreover, friends with dissimilar tastes are treated differently in the social regularization terms in order to represent the taste diversity of each user’s friends. Our proposed approaches are quite general, and they can also be applied to trust-aware recommender systems. The experimental analysis on two large datasets (one dataset contains a social friend network while the other dataset contains a social trust network) shows that our methods outperform other state-of-the-art algorithms.”