Item Cluster-aware Prompt Learning for Session-based Recommendation

Wooseong Yang, Chen Wang, Zihe Song, Weizhi Zhang, Philip S. Yu

Abstract: Session-based recommendation (SBR) captures dynamic user preferences by focusing on item sequences within sessions. However, many existing methods primarily model intra-session item relationships (within a single session) while overlooking inter-session relationships (between items across different sessions), limiting their effectiveness to fully capture complex item relations. Although several studies incorporate inter-session information, they often face high computational complexity, leading to prolonged training times and reduced efficiency. To address these issues, we propose the CLuster-aware Item Prompt learning framework for Session-Based Recommendation (CLIP-SBR). CLIP-SBR consists of two modules: 1) item relationship mining that incorporates a global graph to extract both intra- and inter-session item relationships effectively from session data, and 2) item cluster-aware prompt learning which utilizes soft prompts to integrate the mined item relationships into SBR models efficiently. We validate CLIP-SBR through experiments on eight SBR models and three benchmark datasets. Results show consistent improvements in recommendation performance, demonstrating CLIP-SBR’s potential as a robust framework for SBRs.