Content-only attention Network for Social Recommendation

Bin Wu, Tao Zhang, Yeh-Cheng Chen

With the rapid growth of social Internet technology, social recommender has emerged as a major research hotspot in the recommendation systems. However, traditional graph neural networks does not consider the impact of noise generated by long-distance social relations on recommendation performance. In this work, a content-only multi-relational attention network (CMAN) is proposed for social recommendation. The proposed model owns the following advantages: (i) the comprehensive trust based on the historical interaction records of users and items are integrated into the recursive social dynamic modeling to obtain the comprehensive trust of different users; (ii) social trust information is captured based on the attention network mechanism, so as to solve the problem of weight distribution in the same level domain; (iii) two levels of attention mechanisms are merged into a unified framework to enhance each other. Experiments conducted on two representative datasets demonstrate that the proposed algorithm outperforms previous methods substantially.