Multi-Behavior Sequential Recommendation with Temporal Graph Transformer

Abstract

Modeling time-evolving preferences of users with their sequential item\ninteractions, has attracted increasing attention in many online applications.\nHence, sequential recommender systems have been developed to learn the dynamic\nuser interests from the historical interactions for suggesting items. However,\nthe interaction pattern encoding functions in most existing sequential\nrecommender systems have focused on single type of user-item interactions. In\nmany real-life online platforms, user-item interactive behaviors are often\nmulti-typed (e.g., click, add-to-favorite, purchase) with complex cross-type\nbehavior inter-dependencies. Learning from informative representations of users\nand items based on their multi-typed interaction data, is of great importance\nto accurately characterize the time-evolving user preference. In this work, we\ntackle the dynamic user-item relation learning with the awareness of\nmulti-behavior interactive patterns. Towards this end, we propose a new\nTemporal Graph Transformer (TGT) recommendation framework to jointly capture\ndynamic short-term and long-range user-item interactive patterns, by exploring\nthe evolving correlations across different types of behaviors. The new TGT\nmethod endows the sequential recommendation architecture to distill dedicated\nknowledge for type-specific behavior relational context and the implicit\nbehavior dependencies. Experiments on the real-world datasets indicate that our\nmethod TGT consistently outperforms various state-of-the-art recommendation\nmethods. Our model implementation codes are available at\nhttps://github.com/akaxlh/TGT.\n