Modeling Long & Short-term Interests and Assigning Sample Weight for Multi-behavior Sequential Recommendation

Multi-behavior sequential recommendation aims to predict users' next interested item, by learning dynamic user preferences within their multi-behavior interaction sequences. Users' dynamic preferences are decided by both stable long-term and variable short-term interests. Early efforts towards entangling these two aspects, which may lead to inferior recommendation accuracy and interpretability. Moreover, they ignore the differences in importance between different samples consequently limiting the model-fitting performance. With this concern, we propose a new recommendation framework Multi-_Behavior Interest Matching Network (MB-IMN) to overcome these limitations. Specifically, we model two interest aspects explicitly: a transformer with attention fusion encodes behavior-aware sequential patterns for short-term interest, and a graph learning paradigm is developed to capture multi-behavior interaction semantics for long-term interest. Furthermore, we devise a novel loss function that automatically determines sample importance based on a predefined modulating factor, thus reweighting the samples accordingly. Empirical results on three real-world e-commerce datasets demonstrate the effectiveness of proposed framework. Our implementation code is released at https://github.com/tripiggyo1/MB-IMN.