Source-linked AI summary
Is ChatGPT Fair for Recommendation? Evaluating Fairness in Large Language Model Recommendation
Jizhi Zhang, Keqin Bao, Yang Zhang, Wenjie Wang, Fuli Feng, Xiangnan He
TL;DR
RecLLM raises fairness concerns because LLMs may encode social prejudices, while traditional recommendation benchmarks do not directly fit its open-ended outputs. The paper introduces FaiRLLM, which evaluates user-side fairness across eight sensitive attributes in music and movie recommendation, and finds that ChatGPT generates unfair recommendations for several attributes.
Problem
Evaluating user-side fairness in RecLLM is important because LLMs may reinforce social biases and traditional recommendation fairness benchmarks do not directly fit open-ended outputs.
Method
FaiRLLM uses similarity-based fairness metrics and datasets for music and movie recommendation covering eight sensitive attributes.
Results
ChatGPT generates unfair recommendations across several sensitive attributes in both movie and music recommendation settings.
Takeaways & Limitations
The findings support evaluating fairness when LLMs are used for recommendation.
Takeaways & Limitations
The dataset construction assumes that user instructions follow a fixed template containing preference and task information.
Abstract
from arXiv · showhide
The remarkable achievements of Large Language Models (LLMs) have led to the emergence of a novel recommendation paradigm -- Recommendation via LLM (RecLLM). Nevertheless, it is important to note that LLMs may contain social prejudices, and therefore, the fairness of recommendations made by RecLLM requires further investigation. To avoid the potential risks of RecLLM, it is imperative to evaluate the fairness of RecLLM with respect to various sensitive attributes on the user side. Due to the differences between the RecLLM paradigm and the traditional recommendation paradigm, it is problematic to directly use the fairness benchmark of traditional recommendation. To address the dilemma, we propose a novel benchmark called Fairness of Recommendation via LLM (FaiRLLM). This benchmark comprises carefully crafted metrics and a dataset that accounts for eight sensitive attributes1 in two recommendation scenarios: music and movies. By utilizing our FaiRLLM benchmark, we conducted an evaluation of ChatGPT and discovered that it still exhibits unfairness to some sensitive attributes when generating recommendations. Our code and dataset can be found at https://github.com/jizhi-zhang/FaiRLLM.
1 INTRODUCTION
RecLLM introduces fairness concerns because recommendations may reflect social biases, while traditional fairness benchmarks do not fit its open-ended, instruction-based outputs. The paper proposes FaiRLLM to evaluate user-side fairness and reports unfairness in ChatGPT recommendations.
- RecLLM generates recommendations from natural-language user instructions, creating a new recommendation paradigm.
- LLMs may reinforce social biases, making fairness a critical concern for RecLLM recommendations.
- Traditional fairness benchmarks are ill-suited because RecLLM often lacks prediction scores and produces open-ended recommendation sets.
- FaiRLLM evaluates fairness by comparing neutral-instruction recommendations with sensitive-instruction recommendations and examining similarity differences across attribute values.
- FaiRLLM covers music and movie recommendation datasets with eight sensitive attributes and reveals fairness issues in ChatGPT.
- The paper presents its investigation as the first study of fairness in the emerging LLM recommendation paradigm.
2 RELATED WORK
Prior work studies fairness in LLMs and conventional recommendation systems, but RecLLM requires a distinct user-side group-fairness formulation. The paper emphasizes that its similarity-based fairness target differs from traditional group fairness.
- Research on LLM fairness addresses harmful or discriminatory generation, including methods for mitigation and benchmarks for evaluating unfairness.
- Recommendation fairness research covers both item-side and user-side unfairness and distinguishes individual fairness from group fairness.
- RecLLM fairness is treated as user-side group fairness, but it compares sensitive-group similarity to a neutral group rather than recommendation quality across groups.
3 FAIRLLM BENCHMARK
FaiRLLM is a RecLLM-specific fairness benchmark that compares recommendations under neutral and sensitive instructions across sensitive-attribute values. It combines divergence-based fairness metrics, recommendation-list similarity measures, and datasets for music and movie recommendation.
- Fairness Evaluation: FaiRLLM evaluates user-side fairness by comparing neutral-instruction recommendations with recommendations generated after injecting each sensitive-attribute value.Unfairness is quantified from divergence in similarity across the attribute’s possible values.
- Fairness Metrics: SNSR measures the similarity gap between the most advantaged and disadvantaged groups, while SNSV measures similarity variation across all attribute values.For both metrics, higher values indicate greater unfairness.
- Similarity Metrics: The benchmark computes similarity between neutral and sensitive recommendation lists using Jaccard, SERP*, and PRAG* metrics.Jaccard captures set overlap, SERP* incorporates overlap and item ranks, and PRAG* additionally considers pairwise ranking order.
- Similarity Metrics: PRAG* measures agreement in pairwise ranking between recommendations from neutral and sensitive instructions, not merely item overlap.A higher value requires both substantial overlap and matching relative order between recommended items.
- Dataset Construction: The dataset uses natural-language templates containing preference and task information, with sensitive attributes injected into otherwise neutral instructions.Top-K recommendation lists use item titles as item identities.
- Dataset Construction: The benchmark covers eight sensitive attributes and two recommendation scenarios, using selected singers for music and directors of popular movies and television shows for movies.The attributes are age, country, gender, continent, occupation, race, religion, and physics.
4 RESULTS AND ANALYSIS
The evaluation uses FaiRLLM to examine ChatGPT’s fairness across sensitive attributes and test whether observed unfairness persists under altered list lengths, typos, and languages. ChatGPT shows unfairness across music and movie recommendations, with attribute-specific disparities remaining visible across these conditions.
- The experiments answer whether ChatGPT is unfair across sensitive user attributes and whether unfairness remains robust across different cases.
- Overall Evaluation (RQ1): Table 1 evaluates ChatGPT on music and movie recommendations using SNSR and SNSV, where higher values indicate greater unfairness.Min and Max denote the minimum and maximum similarity across sensitive-attribute values.
- Overall Evaluation (RQ1): ChatGPT demonstrates unfairness across the most sensitive attributes in both movie and music recommendations, with the degree varying by sensitive attribute.The four highest-SNSV attributes are religion, continent, occupation, and country for music, and race, country, continent, and religion for movies.
- The Influence of Recommendation List Length: Similarity differences persist when recommendation lists are truncated to different lengths K, while the relative ordering of sensitive-attribute values remains mostly unchanged.Figure 2 shows the corresponding music results in its top subfigures and movie results in its bottom subfigures.
- The Influence of Sensitive Attribute Typos: Typos closer to the disadvantaged value “African” produce lower similarity to the neutral group, indicating stronger disadvantage.The tested typos include “Afrian” and “Africcan” for “African,” alongside variants of “American.”
- The Influence of Language: Chinese prompts preserve differences among “African,” “American,” and “Asian,” while movie similarities become significantly lower than with English prompts.The lower movie similarity is attributed to outputs mixing Chinese and English recommendation items.
5 CONCLUSION
The paper introduces FaiRLLM to evaluate fairness in generative recommendation across domains and sensitive attributes, then uses it to assess ChatGPT. The experiments find that ChatGPT generates unfair recommendations, while the authors identify evaluation of other models and generative approaches as future work.
- The paper proposes FaiRLLM, combining a fairness evaluation method, specific metrics, and benchmark datasets spanning recommendation domains and eight sensitive attributes.
- Extensive FaiRLLM experiments find that ChatGPT generates unfair recommendations, indicating potential risks in directly applying the RecLLM paradigm.
- The approach is presented as a preliminary attempt to measure fairness in generative text recommendation, with broader generative recommendation settings left for future exploration.