Graph Isomorphic Network with Denoising Attention Mechanism for Session-Based Recommendation
Dingkun Lu, Xiaoang Lu, Zhixin Wang, Jinyu Huang · 2024
The session-based recommendation system aims to predict items that users may interact with in the future by analyzing historical behaviors in anonymous sessions. This field has long faced two major challenges: first, the integration of non-adjacent item relationships is insufficient, which leads to the neglect of higher-order relationship information of items not directly connected; second, the uncertainty in user behavior introduces noise, negatively affecting prediction accuracy. These two issues collectively hinder the recommendation system's ability to accurately capture users' true intentions. To address these problems, this paper proposes a new model named IGANRec. IGANRec learns two-level item embeddings through session graphs and relational graphs, utilizing the powerful aggregation capabilities of Graph Isomorphism Networks (GIN) for initial encoding. Furthermore, by introducing multiple information highways, the model can select effective information before and after GIN and uses a self-attention mechanism to generate target representations. Then, the model performs denoising on the global representation. In the relational graph, a target-aware global attention mechanism is used to generate the denoised relational representation. By integrating information from both graphs and considering the similarity of sessions from the perspective of user intent through the Intent Neighborhood Collaboration Module, the session's intent representation is obtained. Experimental results on three benchmark datasets show that the proposed IGANRec exhibits the improvement of performance compared to existing state-of-the-art models.