Research map / Image generation / Text-to-image diffusion
Intersectional Unfairness Discovery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BGGN, a VAE-based generative model fine-tuned with a bias-value reward (variational/reverse-KL objective, REINFORCE with baseline) to generate diverse high-bias intersectional sensitive attributes on CelebA and Toxic, then uses generated unseen attributes as prompts for LLaMA and Midjourney.
Problem
Existing fairness work mostly examines single sensitive attributes, and enumeration or combinatorial search over intersectional sensitive attributes is exponential and fails to find diverse, unseen high-bias subgroups.
Approach
Model intersectional unfairness discovery as a generative process whose target distribution is proportional to exp(bias value), trained by variational inference with a reverse-KL objective in which the prediction loss on each subgroup acts as a reward (RL-style REINFORCE with baseline), after VAE pre-training.
Key results
- On Toxic at tau=0.3, search tree finds only 4 high-bias attributes with decision-tree estimates while relaxed search (N_re=5) finds 105; BGGN results are reported in figures and are said to discover significantly more high-bias attributes than search baselines.
- BGGN generates a higher share of high-bias attributes than the regular VAE, which mostly produces low-bias samples, per radar charts on CelebA and Toxic (figures not reproduced in text).
- Generated unseen attribute combinations, formulated as prompts, yielded biased outputs from LLaMA, GPT and Midjourney (qualitative findings).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Decision Tree bias estimates with the bias value predictor | Search Tree and Relaxed Search Tree · Toxic, τ=0.3 | discovered high-bias a number: 4 -> 3; 16 -> 16; 61 -> 64; 105 -> 107 | No clear effect |
| Increase the Relaxed Search Tree relaxation number from N_re=1 to N_re=3 or N_re=5 | Relaxed Search Tree · Toxic, τ=0.3 | discovered high-bias a number: 16 -> 61 -> 105 | Helped |
Surprising findings
- Replacing the decision-tree bias estimator with the learned bias predictor barely changes search results (e.g., 4 vs 3 for Search Tree on Toxic).
Limitations
- RL-style training of the generator is unstable, requiring pre-training and resampling tricks.
- Only discovers biased subgroups; does not propose mitigation methods.
- Uses prediction loss as the sole fairness criterion, not demographic parity or equalized odds.
Experimental setup
Datasets: CelebA Toxic Comments
Benchmarks: CelebA Toxic Comments
Metrics: Bias Number Bias Ratio Precision@K Recall@K Avg DCG@K RR@K Score
Baselines: Search Tree Relaxed Search Tree VAE
Builds on: variational autoencoder REINFORCE RLHF
Closest papers by meaning
- Searching for Unfairness in Algorithms’ Outputs: Novel Tests and Insights (AAAI 2025 · similarity 0.88)
- Intersectional Fairness via Mixed-Integer Optimization (NeurIPS 2026 · similarity 0.88)
- AIM-Fair: Advancing Algorithmic Fairness via Selectively Fine-Tuning Biased Models with Contextual Synthetic Data (CVPR 2025 · similarity 0.87)
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.87)
- Hyperbolic Relational Prompts for Intersectional Fairness in Medical VLMs (CVPR 2026 · similarity 0.86)
- Causal Bias Detection in Generative Artifical Intelligence (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.