Fairness identification of large language models in recommendation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Recommender Systems

Background:

• Fairness in recommendation systems is essential for equitable user treatment.
• Large Language Models (LLMs) exhibit human-like behaviors, including fairness awareness.
• Existing recommendation models may inadvertently perpetuate biases.

Purpose of the Study:

• To investigate LLMs as fairness recognizers in recommendation systems.
• To leverage LLMs' fairness awareness for constructing equitable recommendations.
• To propose a method integrating LLMs into the recommendation pipeline.

Main Methods:

• Utilized MovieLens and LastFM datasets for evaluation.
• Compared Variational Autoencoders (VAE) with and without fairness strategies.
• Employed ChatGLM3-6B and Llama2-13B to assess recommendation fairness.
• Developed a hybrid approach using LLMs to refine VAE recommendations.

Main Results:

• LLMs effectively identify unfair recommendations by correlating user attributes with outcomes.
• The proposed method significantly enhances recommendation fairness.
• Minimal loss in recommendation utility was observed post-integration.
• Fairness-to-utility ratios improved dramatically, e.g., from ~5-6 to ~30-50 with ChatGLM.

Conclusions:

• LLMs can serve as effective fairness detectors in recommender systems.
• Integrating LLMs into recommendation processes offers a superior fairness-utility trade-off.
• This approach holds promise for developing more equitable AI-driven systems.