d9d50e6d98f01f357357eafde24ab66370fc3559
Intersectional Fairness Preference Alignment for LLM-based Recommender Systems
LLM-based recommender systems often perpetuate or amplify biases across multiple intersecting demographic attributes, leading to unfair recommendations for marginalized groups. Current approaches focus on single-attribute fairness or use simple demographic parity metrics, failing to capture the compounded effects of multiple overlapping identities.
Existing methods like the CFaiRLLM benchmark evaluate fairness on individual attributes but lack intersectional analysis. By aligning LLM preferences with intersectional fairness criteria, we can create more equitable recommendation systems that account for the complex interplay of multiple demographic factors. Our proposed method, Intersectional Fairness Preference Alignment (IFPA), combines preference learning and intersectional fairness optimization to address these limitations.
IFPA consists of the following steps: 1) Generate diverse synthetic user profiles with intersecting attributes. 2) Create pairwise preference data where the preferred option maximizes an intersectional fairness metric. 3) Fine-tune the LLM using this preference data with Direct Preference Optimization. 4) Implement an intersectional fairness loss during training to encourage equal treatment across all attribute combinations. 5) Apply a fairness-aware decoding strategy at inference time to adjust output probabilities based on intersectional fairness of candidate recommendations.
Step 1: Data Preparation
Generate synthetic user profiles with intersecting attributes (age, gender, race, socioeconomic status) using a probabilistic model. Ensure a balanced distribution across all intersectional groups.
Step 2: Recommendation Datasets
Prepare datasets for multiple domains: movies (MovieLens-1M), books (GoodReads), and job postings (a subset of Indeed job listings). Augment these datasets with synthetic demographic information aligned with the generated user profiles.
Step 3: Baseline Models
Implement standard LLM-based recommenders using GPT-3.5 and GPT-4 APIs. Also implement single-attribute fairness methods as additional baselines.
Step 4: Intersectional Fairness Metric
Implement an intersectional demographic parity metric that measures the difference in recommendation rates across all intersectional groups.
Step 5: Preference Data Generation
For each domain, generate pairwise preference data where the preferred option maximizes the intersectional fairness metric. Use prompts like: 'Given two recommendation lists A and B for a diverse group of users, which list is more fair across all intersectional demographic groups?'
Step 6: LLM Fine-tuning
Fine-tune GPT-3.5 and GPT-4 using the generated preference data with Direct Preference Optimization. Implement the intersectional fairness loss during training.
Step 7: Fairness-aware Decoding
Implement a decoding strategy that adjusts output probabilities based on the intersectional fairness of candidate recommendations. Use a sliding window approach to balance fairness and relevance.
Step 8: Evaluation
Evaluate the IFPA method against baselines on all three domains. Measure recommendation quality (NDCG, MAP) and intersectional fairness metrics. Conduct ablation studies to assess the impact of each component.
Step 9: Analysis
Analyze the trade-offs between recommendation quality and fairness. Examine how IFPA performs across different intersectional groups and identify any remaining biases or limitations.
Baseline Prompt Input
Recommend 5 movies for a 25-year-old Black woman from a low-income background who enjoys science fiction and drama.
Baseline Prompt Expected Output
Proposed Prompt Input
Using the Intersectional Fairness Preference Alignment method, recommend 5 movies for a 25-year-old Black woman from a low-income background who enjoys science fiction and drama. Ensure the recommendations are fair across all intersectional demographic groups.
Proposed Prompt Expected Output
Explanation
The IFPA method provides a more diverse set of recommendations that better represent the user's intersectional identity while still aligning with their genre preferences. It includes films with Black leads and themes relevant to the user's background, balancing mainstream sci-fi with more diverse storytelling.
If IFPA doesn't significantly improve intersectional fairness, we can pivot to an in-depth analysis of why LLMs struggle with intersectional fairness in recommendations. We'll examine the generated preference data, fine-tuning process, and decoding strategy to identify where biases are introduced or amplified. We can also explore alternative fairness metrics or multi-objective optimization approaches that balance individual and group fairness. Additionally, we might investigate how different prompting strategies affect intersectional fairness, potentially leading to insights on how to better elicit fair behavior from LLMs without extensive fine-tuning.