Multi-behavior Guided Temporal Graph Attention Network for Recommendation

Abstract

The traditional recommendation approaches learn the representations of users and items utilizing only a single type of behavior data, which results in them facing the data sparsity issue. To alleviate the dilemma, multi-behavior recommendations leverage different types of behaviors to assist in modeling users’ preferences. Despite their remarkable effectiveness, two significant challenges have remained less explored: 1) Effectively distinguishing the contributions of different types of behaviors during capturing users’ preferences; 2) Sufficiently exploiting the temporal information of user-item interactions. To tackle these challenges, we develop a new model named Multi-behavior Guided Temporal Graph Attention Network (MB-TGAT) to discriminate the diverse influence of various behaviors and to explore the evolutionary tendencies of users’ recent preferences. In particular, we propose a behavior-aware attention mechanism to differentiate the strengths of different behaviors in the user-item aggregation phase. Furthermore, we tailor a phased message passing mechanism based on GNNs and design an evolution sequence self-attention to extract the users’ preferences from static and dynamic perspectives, respectively. Extensive experiments on three real-world datasets demonstrate the superiority of our model, noticeably with 37.27%, 37.31% and 14.63% performance gain over the state-of-the-art baselines on the Taobao, IJCAI-15 and YooChoose datasets, respectively.