Improving session-based recommendation with contrastive learning

Session-based recommendation, which aims to predict the next item given anonymous behavior sequences of users, is critical in modern recommender systems. While prior works have made efforts to improve recommendation performance, two challenges remain unsolved. First, existing learning methodologies rely on mining sequential patterns within the individual session and use the next item as the supervised signal, which may not effectively capture the correlations among interactions. Second, previous solutions are also limited in learning the mixed dependencies inside flexibly ordered sessions, i.e., sequential dependencies among ordered interactions and non-sequential dependencies among unordered ones. This work presents a novel session recommender algorithm by distilling knowledge and supervision signals from sessions in a contrastive manner. We propose position-aware importance extraction module with contrastive learning, which utilizes the intrinsic dependencies to discover extra knowledge and augment the ability of information distillation. Besides, we introduce a bi-directional matching algorithm with contrastive loss to capture potential patterns through maximizing the mutual information between current interaction and historical behaviors. Moreover, we introduce a simple yet effective learnable position-coding mechanism with self-attention-based importance extraction to flexibly learn user browsing patterns. Extensive experiments conducted on two real-world datasets demonstrate that our proposed algorithm enhances the recommendation performance over existing state-of-the-art approaches.