Cross-view temporal graph contrastive learning for session-based recommendation

Abstract

Session-based recommendation (SBR) aims at recommending items given the behavior sequences of anonymous users in a short-term session. Many recent SBR methods construct all sessions as a global graph that captures cross-session item transition patterns (i.e., users’ global preference) to alleviate the problem of session data sparsity. However, these methods neglect that users’ interests will drift as the time between the sessions constantly increases, limiting the performance improvement of SBR. To fill this gap, we divide all sessions into a group of time-slices and model the cross-session item transitions for every time-slice. We further construct two augmentation views (i.e., temporal graph and temporal hypergraph views) to model pairwise and high-order item transitions on SBR. In addition, we construct contrastive learning between two views to improve the recommendation performance by maximizing the mutual information between the item representations learned from the two views. Experiments on three public real-world datasets (i.e., Diginetica, Retailrocket, and Yoochoose) show that our model is consistently superior to the other state-of-the-art baselines, especially in time-sensitive datasets. For instance, our model achieves 14.42% and 11.72% improvements in terms of P@10 and P@20 on the Diginetica dataset, respectively.