ACCF: Learning Attentional Conformity for Collaborative Filtering

In recent years, Collaborative Filtering (CF) methods have yielded immense success on recommender systems. They mainly use the similarity between users and items, or the interactions between users and items to predict the unknown ratings. However, the social conformity phenomenon received little notice, which means: 1) individuals in a social network can have multiple characteristics and hence tend to belong to multiple overlapping groups or communities and 2) when confronted with conformity pressure, people often adjust their responses to conform to others’ opinions to obtain social approval and belonging in the community. In this paper, we propose a new collaborative filtering-based recommendation framework, called ACCF, which explicitly exploits social conformity of users. We incorporate such social conformity phenomenon into the latent factor model, using a weighted average of community preference profiles as the adjusting factor, and learn the weight of each community’s influence through an attention network. Compared with the seven state-of-the-art methods on three real-world datasets, our method achieves the best performance.