Fairness Retrieval

This research studies exposure fairness in the retrieval stage of recommender systems. Accuracy-oriented recommendation models may repeatedly expose popular items or dominant item groups, leaving other groups underrepresented in the candidate set. To address this issue, this work proposes PAFER, a preference-aware fairness exposure retrieval framework that improves item-group exposure while considering each user’s group-level preference structure.

PAFER estimates user-group affinity and constructs user-specific masks so that exposure compensation is applied only to groups that are compatible with the current user. It further adopts an entropy-aware candidate construction strategy: users with broader interests receive more flexible group-level exploration, while users with concentrated interests are handled more conservatively to preserve recommendation quality and preference consistency.