V. V. Ivanova, Evgeny Frolov · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840006
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).
Sequential recommender systems are commonly optimized for next-item prediction, but they often underutilize negative feedback such as dislikes, skips, or low ratings. As a result, items that a user is likely to reject may still be ranked highly, which can reduce user satisfaction. Effectively incorporating negative feedback is challenging because such signals are often sparse, noisy, and must be modeled without degrading positive-preference prediction. We propose NegBT, a non-contrastive framework for negative-feedback-aware sequential recommendation based on the Barlow Twins objective. The key idea is to construct two deterministic views of the same user's history: a positive-only sequence that captures core preferences and a full mixed-feedback sequence that preserves the original behavioral context. By encouraging consistency between these representations while reducing feature redundancy, NegBT incorporates mixed-feedback histories as an auxiliary regularizer without relying on hard negative sampling. Experiments on three datasets show that NegBT provides a strong balance between positive-item accuracy and negative-feedback avoidance. In particular, it improves over positive-only SASRec on positive-target ranking metrics while reducing the relative promotion of negative items, with especially clear gains on explicit-feedback datasets and long user histories.
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.