Hierarchical Session-Aware Recommendation Model Based on Temporal Reasoning

Session-Based Recommendation Systems (SRS) aim to predict users' next behaviors via anonymous sessions, with the core challenge of simulating the temporal relationships of user behaviors and depicting user interests using limited sessions. Existing methods mostly model behavior patterns based on the temporal relationships of adjacent items in sessions and selectively aggregate item information as user interests. This paper proposes a Temporal Reasoning-based Hierarchical Session-Aware Recommendation Model (TRHSR): on one hand, it breaks the assumption of “adjacent items being relevant”, infers dependencies between items to learn more flexible temporal relationships; on the other hand, it aggregates information at both the item and item feature levels to achieve fine-grained interest inference. Experiments on two public datasets (Diginetica and Retailrocket) show that the model outperforms other benchmark models, verifying its effectiveness.