Contrastive learning enhanced multi-level interest-aware for session-based recommendation via graph attention networks
Fan Yang, Dunlu Peng, Yiming Xu, Nan Chen · Expert Systems with Applications · 2026
The objective of session-based recommendation (SBR) is to predict the next item by leveraging a sequence of interactions. Accurately modeling the dependency of the user’s next click on the past action is crucial to improve the recommendation performance. Existing SBR models focus heavily on the user’s current interest represented by the most recent item in the sequence or the interaction between single items, without fully considering different levels of user’s interests. In addition, the sparsity of interaction sequence also limits the recommendation performance of the model. This work proposes contrastive learning enhanced multi-level interest-aware graph attention networks (CLEMI-GAT) for SBR. The session sequence is constructed as a multi-level interest-aware hypergraph, which utilizes interest-aware modules at different levels and hypergraph attention networks (HGAT) to learn deep interests of users. Simultaneously, a interaction graph is built to aggregate the long-term attention of users, and the intermediary node is added to acquire long-distance dependencies among items. A fusion gated network is employed to merge multi-level interests and long-term attention of users. In this work, repeat-exploration normalization (RENorm) is introduced to account for both exploration and repetition behaviors in predictions, which is to better distinguish items within and outside the session sequence. The auxiliary task, namely contrastive learning (CL), further enhances the item representation of our model CLEMI-GAT, thereby alleviating data sparsity. Extensive experimental results show that CLEMI-GAT is superior to the baseline models on the realistic datasets.