Time-based Knowledge-aware framework for Multi-Behavior Recommendation

Multi-behavior recommendation alleviates the data sparsity problem in single-behavior recommendation by exploiting the multi-dimensional behavioral information of users to construct rich connections between users and items. However, existing multi-behavior recommendation methods tend to ignore the temporal information of user interactions, which makes it difficult to dynamically understand user preferences. In addition, introducing too much behavioral information may lead to negative migration problem in the model (i.e., the newly introduced behavioral information conflicts with the original information), which leads to model performance degradation. Based on the above background and challenges, we propose a Time-based Knowledge-aware Multi-Behavior Recommendation framework (TKMB). The framework combines multi-behavior and temporal information of users, and achieves comprehensive modeling of user preferences and item information through three main views: the local multi-behavior interaction view, the global multi-behavior interaction view and the knowledge-aware view. The first two separately design a local and global self-attention mechanism to distinguish the importance of different behaviors. And designs an adaptive time gating mechanism to dynamically capture users’ personalized preferences. The latter constructs high-order representations at the item level and proposes a graph reconstruction strategy and knowledge-aware contrastive learning to enhance the robustness of the model. Finally, a multi-view aggregation mechanism is introduced to aggregate multi-scale representations. The results of extensive experiments and ablation experiments on two real datasets further validate the effectiveness and superiority of TKMB.