Yuhang Zhang · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840899
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).
In sequential recommendation, multi-interest models are widely deployed to capture diverse user preferences from behavioral histories. However, existing architectures typically extract multiple representation vectors and fuse them via conventional attention mechanisms. By computing a weighted sum, this paradigm implicitly assumes that distinct preferences are independent and can always be linearly combined. Consequently, when internal conflicts arise, such as budget constraints versus performance requirements, these feed-forward mechanisms compromise, yielding recommendations that combine incompatible signals. In this paper, we formulate conflicting preferences through explicit pairwise constraints. We propose Constrained Preference Dynamics (CPD), a framework that casts preference aggregation as a constrained dynamical system. CPD treats preference intensities as state variables evolving under learned pairwise constraints. By minimizing an energy function via unrolled gradient descent, the system moves preference states toward locally stable, low-energy configurations before scoring. On five real-world datasets, CPD improves overall accuracy while delivering larger gains under high-conflict scenarios. Specifically, on the Amazon Clothing conflict subset, CPD outperforms the strongest baseline MDGR by 19.8% in Recall@10 and 24.8% in NDCG@10. These results show that incorporating constrained dynamics improves recommendation performance under preference conflict.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
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.