Improving Recommendation Fairness without Sensitive Attributes Using Multi-Persona LLMs

TL;DR

This paper tackles fair recommender systems under no access to sensitive attributes by introducing LLMFOSA, an LLM-enhanced framework that uses multi-persona sensitive information inference and confusion-aware representation learning. A two-stage mutual-information optimization guides the learning of sensitive-aware and sensitive-blind embeddings, enabling fair recommendations without explicit demographic data. The approach demonstrates substantial fairness improvements on MovieLens-1M and LastFM-360K with modest accuracy trade-offs, outperforming baselines that rely on sensitive attributes or no attribute information. By leveraging external knowledge from large language models and robust MI-based objectives, the method offers a privacy-conscious, scalable pathway to more equitable recommendations in real-world platforms.

Abstract

Despite the success of recommender systems in alleviating information overload, fairness issues have raised concerns in recent years, potentially leading to unequal treatment for certain user groups. While efforts have been made to improve recommendation fairness, they often assume that users' sensitive attributes are available during model training. However, collecting sensitive information can be difficult, especially on platforms that involve no personal information disclosure. Therefore, we aim to improve recommendation fairness without any access to sensitive attributes. However, this is a non-trivial task because uncovering latent sensitive patterns from complicated user behaviors without explicit sensitive attributes can be difficult. Consequently, suboptimal estimates of sensitive distributions can hinder the fairness training process. To address these challenges, leveraging the remarkable reasoning abilities of Large Language Models (LLMs), we propose a novel LLM-enhanced framework for Fair recommendation withOut Sensitive Attributes (LLMFOSA). A Multi-Persona Sensitive Information Inference module employs LLMs with distinct personas that mimic diverse human perceptions to infer and distill sensitive information. Furthermore, a Confusion-Aware Sensitive Representation Learning module incorporates inference results and rationales to develop robust sensitive representations, considering the mislabeling confusion and collective consensus among agents. The model is then optimized by a formulated mutual information objective. Extensive experiments on two public datasets validate the effectiveness of LLMFOSA in improving fairness.

Figures (9)

• Figure 1: The overview of the LLMFOSA framework. 1) LLM-powered agents collaboratively infer users' sensitive information through diversified roles. 2) Inference results and rationales are incorporated, considering mislabeling confusion and collective consensus to develop robust sensitive representations. 3) Fairness is achieved by optimizing the formulated MI objective.
• Figure 2: Contributions of different components in LLMFOSA to the fairness.
• Figure 3: The performance on group fairness metrics.
• Figure 4: Parameter sensitivity on two hyper-parameters. Results of accuracy are in \ref{['app:ablation_sensitivity']}.
• Figure 5: Contributions of different components in LLMFOSA to the recommendation accuracy.