[Paper Review] Is ChatGPT Fair for Recommendation? Evaluating Fairness in Large Language Model Recommendation

Jizhi Zhang, Keqin Bao+4 more|arXiv (Cornell University)|May 12, 2023

Topic Modeling41 references17 citations

TL;DR

The paper introduces FaiRLLM, a fairness benchmark for RecLLM, and assesses ChatGPT, revealing unfairness across several user-sensitive attributes in music and movie recommendations.

ABSTRACT

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.

Motivation & Objective

• Motivate fairness concerns in the RecLLM paradigm where LLMs generate recommendations based on user instructions.
• Propose FaiRLLM, a benchmark with metrics and datasets for eight sensitive user attributes in two domains (music, movies).
• Define a neutral-versus-sensitive instruction similarity framework to assess fairness without requiring candidate scores.
• Provide reproducible evaluation results using ChatGPT to reveal biases and suggest directions for mitigation.

Proposed method

• Define a fairness notion for RecLLM as the absence of prejudice toward user groups when sensitive attributes are not provided.
• Compute top-K recommendation similarities between neutral instructions and sensitive-instruction variations.
• Introduce three similarity metrics (Jaccard@K, SERP*@K, PRAG*@K) to compare neutral and sensitive lists while accounting for ranking.
• Develop two fairness metrics (SNSR@K, SNSV@K) that quantify divergence of aggregated similarities across sensitive attribute values.
• Construct two benchmark datasets (Music, Movie) with eight sensitive attribute values and a template-based instruction format.
• Use a greedy decoding setup for ChatGPT to ensure reproducibility in generating top-K lists.

Experimental results

Research questions

• RQ1 RQ1: How unfair is the LLM when serving as a recommender across various sensitive user attributes?
• RQ2 RQ2: Is the observed unfairness robust across conditions such as prompt typos or different languages?
• RQ3 RQ3: Do the fairness measurements reflect consistent patterns across music and movie domains?

Key findings

• ChatGPT exhibits unfairness across multiple sensitive attributes in both music and movie recommendations.
• Fairness metrics SNSR and SNSV reveal attribute-dependent unfairness levels; religion, continent, occupation, country, race, and others show varying impact.
• Unfairness persists when truncating recommendation lists to different lengths (K).
• Typos in sensitive attribute values can worsen or preserve unfairness, depending on proximity to disadvantaged groups.
• Chinese prompts also show persistent unfairness, with some domains (movie) showing lower similarity under translation scenarios.