DINE: Dynamic Information Network Embedding for Social Recommendation

Abstract

The social recommendation aims to integrate social network information to improve the accuracy of traditional recommender systems. Learning embeddings of nodes in networks is one of the core problems of many applications such as recommendation, link prediction, and node classification. Early studies cast the social recommendation task as a vertex ranking problem. Although these methods are effective to some extent, they require assuming social networks and user-item interaction networks as static graphs, whereas real-world information networks evolve over time. In addition, the existing works have primarily focused on modeling users in social networks in general and overlooked the special properties of items. To address these issues, we propose a new method named DINE, short for Dynamic Information Network Embedding, to learn the vertex representations for dynamic networks in social recommendation task. We model both users and items simultaneously and integrate the representations in dynamic and static information networks. In addition, the multi-head self-attention mechanism is employed to model the evolution patterns of dynamic information networks from multiple perspectives. We conduct extensive experiments on Ciao and Epinions datasets. Both quantitative results and qualitative analysis verify the effectiveness and rationality of our DINE method.

Y. Zhang and D. Meng—These authors contribute equally to this work.