Yu Qing, Xiao Wang, Zhe Wang, Xiurui Xie, Ke Qin, Shuang Liang · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840905
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Session-based Recommendation (SBR) aims to predict users' next interaction based on short-term anonymous behavior sequences. Yet accurately modeling the dynamic switching and hierarchical dependency of user intent remains a critical challenge. Existing methods typically treat intents as parallel signals thus ignoring the dynamic shift in the dominant decision-making pattern layer within a session. However, we argue that the dominant decision-making pattern shifts across hierarchical intent layers over time, and treating these layers as static or independent inevitably causes mismodeling. To address these issues, we propose a Pyramid Routing for Intent Sequence Modeling framework (PRISM), which constructs an intent pyramid to parse user decision-making processes. Specifically, we first devise a Hierarchical Vector Quantization module to progressively compress actions into theme and goals. We then employ a Multi-Intent Routing Network to model the cross-layer dependency and information propagation among different intent layers. Finally, to adapt to the dynamic shifting of decision dominance, we design a Mixture of Intent Module that adaptively weights the prediction distributions from different layers, and introduce a Goal-Aware Contrastive Regularizer to ensure the consistency of hierarchical intent representations. Experimental results on three real-world datasets exhibit that PRISM achieves state-of-the-art performance.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.