Relational metric learning with high-order neighborhood interactions for social recommendation

Social information has been widely incorporated into traditional recommendation systems to alleviate the data sparsity and cold-start issues. However, existing social recommend methods typically have two common limitations: (a) they learn a unified representation for each user involved in both item and social domains, which is insufficient for fine-grained user modeling. (b) They ignore the high-order neighborhood information encoded in both user-item interactions and social relations. To overcome these two limitations, this paper proposes a novel social recommend method, SoHRML, based on social relations under the metric learning framework. Specifically, user-item interactions and social relations are modeled as two types of relation vectors, with which each user could be translated to both multiple item-aware and social-aware representations. In addition, to capture the rich information encoded in local neighborhoods, we model the relation vectors by high-order neighborhood interactions (HNI). In each domain, we design a dual layer-wise neighborhood aggregation (LNA) structure that contains dual graph attention networks (GATs) to aggregate the neighborhoods of users or items. Both high-order information encoded in user-item interactions and social relations can be captured by stacking the layer-wise structure. Extensive experimental results on three practical datasets demonstrate the superiority of the proposed model, especially under the cold-start scenarios. The performance gains over the best baseline are 0.51% to 3.31% on two ranking-based metrics.