Investigating and Mitigating Stereotype-aware Unfairness in LLM-based Recommendations

TL;DR

This work addresses stereotype-aware unfairness in LLM-based recommender systems by introducing a two-sided fairness framework that accounts for biases in both user and item representations. It proposes Mixture-of-Stereotypes (MoS), a multi-expert architecture that uses multi-stereotype prompting, stereotype-wise routing, and a set of learning objectives to learn unbiased, stereotype-aware representations while preserving recommendation quality. The authors formalize a calibrated, two-sided fairness metric and demonstrate, through substantial experiments on real-world datasets, that MoS outperforms single-sided fairness baselines and conventional LLM-RS paradigms in fairness with competitive or better accuracy. These findings advance the trustworthiness of LLM-RS and offer a scalable approach to mitigating intrinsic stereotypes embedded in textual user/item information, with potential downstream impacts on fairness-aware personalization in real systems.

Abstract

Large Language Models (LLMs) have demonstrated unprecedented language understanding and reasoning capabilities to capture diverse user preferences and advance personalized recommendations. Despite the growing interest in LLM-based recommendations, unique challenges are brought to the trustworthiness of LLM-based recommender systems (LLM-RS). Compared to unique user/item representations in conventional recommender systems, users and items share the textual representation (e.g., word embeddings) in LLM-based recommendations. Recent studies have revealed that LLMs are likely to inherit stereotypes that are embedded ubiquitously in word embeddings, due to their training on large-scale uncurated datasets. This leads to LLM-RS exhibiting stereotypical linguistic associations between users and items, causing a form of two-sided (i.e., user-to-item) recommendation fairness. However, there remains a lack of studies investigating the unfairness of LLM-RS due to intrinsic stereotypes, which can simultaneously involve user and item groups. To bridge this gap, this study reveals a new variant of fairness between stereotype groups containing both users and items, to quantify discrimination against stereotypes in LLM-RS. Moreover, in this paper, to mitigate stereotype-aware unfairness in textual user and item representations, we propose a novel framework named Mixture-of-Stereotypes (MoS). In particular, an insightful stereotype-wise routing strategy over multiple stereotype-relevant experts is designed, aiming to learn unbiased representations against different stereotypes in LLM-RS. Extensive experiments are conducted to analyze the influence of stereotype-aware fairness in LLM-RS and the effectiveness of our proposed methods, which consistently outperform competitive benchmarks under various fairness settings.

Figures (6)

• Figure 1: Illustration of stereotype groups. Compared to unique user/item representations in conventional recommender systems, users and items share the textual representation (e.g., word embeddings) in LLM-based recommendations. Therefore, LLM stereotypes can simultaneously involve user and item groups, leading to a form of two-sided recommendation fairness.
• Figure 2: Illustration of stereotype-aware fairness. In A and B, existing user-side or item-side fairness still ignores two-sided (i.e., user-to-item) stereotypes, only considering the fairness against single-sided stereotypes. In C, we model the fairness against two-sided (i.e., user-to-item) stereotypes based on calibrated proportions. For example, the user-side stereotype can unfairly amplify the original distribution of item-side stereotypes, by over-recommending items (i.e., increase from 70% to 95%) in the female stereotype group and under-recommending items (i.e., decrease from 30% to 5%) in the male stereotype group.
• Figure 3: Threshold of $d_{v \in G}$ based on Z-scores in different experimental datasets. An ablation study is conducted to validate the above settings of threshold in Fig. \ref{['fig:ablation_item']}.
• Figure 4: The overall framework of the proposed MoS. In (a), multi-stereotype prompting elicits biases with respect to different stereotype groups. In (b), MoS mitigates the elicited stereotypes in recommendation tasks, where unbiased representations are generated and integrated with LLM-RS via soft prompts. In (c), effective learning objectives are designed to facilitate both the recommendation performance and the stereotype-aware fairness.
• Figure 5: Results of ablation studies on user-side and item-side stereotypes. The detailed statistics of users $u$, items $v$, and groups $G_1, G_2$ can be found in Table \ref{['tab:dataset']}.