Fairness in Recommender Systems

Introduction

Recommender systems play an important role in helping users discover relevant items from large-scale item collections. However, when a recommender system optimizes only for accuracy, it may repeatedly recommend popular items or dominant item groups, causing some item groups to receive insufficient exposure over time. This problem is especially important in two-stage recommendation pipelines, where the retrieval stage first selects a small candidate set before the ranking stage performs fine-grained ordering. If some item groups are rarely retrieved in the first stage, the later ranking stage has limited ability to recover their exposure.

This research focuses on exposure fairness in the retrieval stage of recommender systems. Instead of applying the same exposure compensation strategy to all users, we argue that fairness-aware retrieval should consider each user's group-level preference structure. Some users have broad interests across multiple item groups, while others have more concentrated preferences. Therefore, exposure adjustment should be applied in a preference-aware manner, so that underexposed groups can receive more opportunities without strongly deviating from the user's original interests.

To address this problem, we propose Preference-Aware Fairness Exposure Retrieval (PAFER), a retrieval-stage framework that improves item-group exposure fairness while preserving recommendation relevance, group diversity, and preference consistency.

The architecture of PAFER

Method

PAFER consists of four main components.

First, PAFER learns group embeddings from item embeddings. Since item groups are originally represented only by category labels, the method transforms each group into a dense representation in the same latent space as users and items. This allows the system to estimate the compatibility between each user and each item group.

Second, PAFER estimates user-group affinity and constructs a user-specific group mask. The affinity distribution describes how strongly a user is related to different item groups. Based on this distribution, PAFER activates only the groups that are compatible with the current user. The mask does not remove items from the retrieval corpus; instead, it controls which groups can receive exposure compensation for each user.

Third, PAFER applies masked exposure compensation in the retrieval stage. Built on the FairSync retrieval framework, PAFER modifies the group-level exposure compensation vector with the user-specific mask. This allows the model to improve item-group exposure while avoiding uniform compensation over all users and all groups.

Finally, PAFER uses an entropy-aware candidate list construction strategy. Users with broader group-level interests are treated as high-entropy users and receive a relevance-preserving head with a group-aware tail, allowing controlled exploration of compatible groups. Users with concentrated interests are treated more conservatively, so that the final candidate list remains close to their original preference structure.

Through these components, PAFER aims to balance four objectives: retrieval effectiveness, exposure fairness, group diversity, and preference consistency. The goal is not only to satisfy item-group exposure requirements, but also to ensure that the retrieved candidate lists remain relevant and compatible with user preferences.

Experiment Result