Cross-Spectrum Candidate Set Analysis for Popularity-Aware Recommendation

초록

Popularity-biased recommender systems can produce unequal recommendation utility across users with different popularity preferences. In particular, users who prefer niche items can experience lower utility even when the system's overall performance improves. To examine this disparity, we propose a multi-task framework with group-specialized heads and perform cross-spectrum candidate set analysis to characterize user preferences. By analyzing intersections among their candidate sets, we can identify group-specific and cross-group consensus items. Experiments reveal varying candidate overlap structures and popularity patterns across models and datasets, suggesting that their set intersections may support future popularity-aware reranking. The source code is available at https://github.com/hddni/Cross-Spectrum-Candidate-Set-Analysis-for-Popularity-Aware-Recommendation. © 2026 Copyright held by the owner/author(s).

키워드