Bi-View Contrastive Learning with Hypergraph for Enhanced Session-Based Recommendation

Session-based recommendation (SBR) aims to predict a user’s next interests based on their actions in a single visit. Recent methods utilize graph neural networks to study the pairwise relationship of item transfers, yet these often overlook the complex high-order connections between items. Hypergraphs can naturally model many-to-many relationships and capture complex interactions, thereby improving the accuracy of SBR. However, the potential of hypergraphs in SBR remains underexplored. This paper models session data into two views: the hypergraph view, which employs hypergraph convolution, and the session view, which utilizes the intersection graph of the hypergraph with standard graph convolution to support the main recommendation task. By combining cross-view contrastive learning with view adversarial training as an auxiliary task, the two views recursively exploit different connectivity information to generate ground truth samples, thus enriching the session information. Extensive experiments on three benchmark datasets confirm the effectiveness of our hypergraph modeling approach and cross-view contrastive learning.