Chen Yi Xu, Hengyu Zhang, Hong Cheng · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840052
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).
Popularity bias is common in sequential recommenders, disproportionately favoring frequent items over long-tail ones. We identify representation anisotropy as a geometric manifestation of this bias: item embeddings collapse toward a cone whose dominant direction correlates with item frequency. To address this, we propose Orthogonal Popularity Decomposition (OPD), which introduces a learnable popularity direction and projects item embeddings onto it, isolating a popularity component from an orthogonal collaborative subspace. OPD produces a near-isotropic embedding subspace and provides an interpretable, controllable popularity prior that can be tuned at inference without retraining. Experiments on three benchmark datasets demonstrate that OPD enables controllable accuracy–fairness trade-offs and consistent performance gains across all datasets. The source code is available online.
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.