Enhancing Session-Based Recommendation With Multi-Interest Hyperbolic Representation Networks.

Abstract

Session-based recommendation (SBR) aims to predict the next item a user might click within an ongoing session, without relying on user profiles or historical data. Modern approaches typically use graph networks to learn item embeddings in Euclidean space via graph convolution operations. However, they often struggle to capture the diversity of user interactions within short, hierarchically structured sessions, which is essential for accurate predictions in SBR. To tackle these challenges, we propose a multi-interest hyperbolic representation network (MIHRN) to enhance the performance of SBR by adeptly modeling both intricate high-order spatial structures and sequence relationships among items in hyperbolic geometry space. Specifically, we use a hyperbolic hypergraph neural network to exploit the high-order spatial relationships and local clustering structures inherent within sessions. Subsequently, a multiaspect interest representation module is designed to articulate the diversity of user interests. Extensive experiments on three real-world datasets demonstrate that the proposed method achieves performance improvements of 23.81%, 14.81%, and 36.84%, respectively, under the P@10 metric.