Social Collaborative Mutual Learning for Item Recommendation

Recommender Systems (RSs) provide users with item choices based on their preferences reflected in past interactions and become important tools to alleviate the information overload problem for users. However, in real-world scenarios, the user-item interaction matrix is generally sparse, leading to the poor performance of recommendation methods. To cope with this problem, social information is introduced into these methods in several ways, such as regularization, ensemble, and sampling. However, these strategies to use social information have their limitations. The regularization and ensemble strategies may suffer from the over-smoothing problem, while the sampling-based strategy may be affected by the overfitting problem. To overcome the limitations of the previous efforts, a novel social recommendation model, namely, Social Collaborative Mutual Learning (SCML), is proposed in this article. SCML combines the item-based CF model with the social CF model by two well-designed mutual regularization strategies. The embedding-level mutual regularization forces the user representations in two models to be close, and the output-level mutual regularization matches the distributions of the predictions in two models. Extensive experiments on three public datasets show that SCML significantly outperforms the baseline methods and the proposed mutual regularization strategies can embrace the advantages of the item-based CF model and the social CF model to improve the recommendation performance.