Relational Metric Learning with Dual Graph Attention Networks for Social Recommendation

Existing social recommenders typically incorporate all social relations into user preference modeling, while social connections are not always built on common interests. In addition, they often learn a single vector for each user involved in two domains, which is insufficient to reveal user's complex interests to both items and friends. To tackle the above issues, in this paper, we consider modeling the user-item interactions and social relations simultaneously and propose a novel metric learning-based model called RML-DGATs. Specifically, relations in two domains are modeled as two types of relation vectors, with which each user can be regarded as being translated to both multiple item-aware and social-aware representations. Then we model the relation vectors by neighborhood interactions with two carefully designed dual GATs to fully encode the neighborhood information. Finally, the two parts are jointly trained under a dual metric learning framework. Extensive experiments on two real-world datasets demonstrate that our model outperforms the best baseline by 1.91% to 4.74% on three metrics for top-N recommendation and the performance gains are more significant under the cold-start scenarios.