Adaptive social similarities for recommender systems

Collaborative filtering (CF) is an effective recommendation technique, which selects items for an individual user based on similar users' preferences. However, CF may not fully reflect the procedure how people choose an item in real life, for users are more likely to ask friends for opinions instead of asking similar strangers. Recently, some recommendation methods based on social network have been raised. These approaches incorporate social network into the CF algorithms and users' preferences can be influenced by the favors of their friends. These social approaches require the knowledge of similarities among friends. There are two popular similarity functions: Vector Space Similarity (VSS) and Pearson Correlation Coefficient (PCC). However, both friends similarity functions are based on the item-sets they rated in common. In most cases, these functions are impractical, i.e. if two friends do not share the same items in common, the similarity between them will be zeros. To solve this problem, we propose an Adaptive Social Similarity (ASS) function based on the matrix factorization technique. We conduct our experiment on a large dataset: Epinions, which is a widely-used dataset with social information. The experiment results illustrate that our approach outperforms the baseline models and achieves a better performance than social-based method in 4.