Research map / Learning theory and algorithms / Fairness
Multi-Group Proportional Representation in Retrieval
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.
Introduces Multi-Group Proportional Representation (MPR), a worst-case metric over a function class of intersectional groups, with estimation guarantees and the MOPR cutting-plane retrieval algorithm that trades off similarity against MPR on face image retrieval.
Problem
Image retrieval can under-represent intersectional demographic groups (e.g., gender combined with race), and existing equal or proportional representation methods only handle a small number of pre-defined binary groups.
Approach
Define MPR as the supremum over a rich function class of deviations between the retrieved-set average and the reference population expectation, show it reduces to an MSE regression oracle (or closed form for linear classes), and use a cutting-plane LP loop (MOPR) to maximize query similarity subject to an MPR constraint.
Key results
- MOPR Pareto-dominates baselines (DebiasCLIP, CLIP-Clip, PBM, MMR) in the similarity-MPR tradeoff on CelebA, UTKFace, and Occupations with k=50 over 10 queries.
- With a balanced synthetic curated dataset, MOPR achieves equal representation across race-gender intersectional groups on UTKFace (each group at 20% sum), while top-k and MMR leave several intersectional groups underrepresented.
- MOPR can drive MPR to near zero on UTKFace, while gaps remain on CelebA and Occupations, which the authors attribute to probe-predicted attribute labels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize retrieval for MPR using an equally balanced curated dataset | MOPR · UTKFace, equal representation across race and gender groups | representation balance | Helped |
| Use MOPR instead of top-k retrieval | CLIP retrieval · UTKFace, equal representation across race and gender groups | representation balance: Top-k group sums include 38.4 ± 17.18 (White) and 3.0 ± 2.04 (Others); MOPR reports 20 ± 0 for each group | Helped |
| Use MOPR instead of MMR | CLIP retrieval · UTKFace, equal representation across race and gender groups | representation balance: MMR group sums include 45.0 ± 7.76 (White) and 6.4 ± 1.96 (Others); MOPR reports 20 ± 0 for each group | Helped |
| Optimize retrieval for MPR while preserving similarity | MOPR · CelebA, UTKFace, and Occupations retrieval | MPR gap and cosine similarity | Mixed |
Surprising findings
- Optimizing for equal or proportional representation over individual attributes does not guarantee representation across intersectional groups.
- Existing fair-retrieval methods that modify embeddings for a single attribute only partially reduce MPR at a high utility cost.
Limitations
- MPR is bounded by the quality and bias of the curated dataset, which propagates to results.
- Cutting-plane algorithm complexity may be a concern for high-dimensional feature spaces and general function classes.
- Group labels for CelebA and Occupations are predicted by linear probes, which may introduce error and ethical concerns.
Experimental setup
Datasets: CelebA UTKFace Occupations FairFace
Benchmarks: CelebA UTKFace Occupations
Metrics: MPR average cosine similarity percentage representation per intersectional group
Baselines: DebiasCLIP CLIP-Clip PBM (Post-Hoc Bias Mitigation) MMR k-NN
Builds on: Maximal Marginal Relevance multi-group fairness multi-attribute proportional representation PATHS
Closest papers by meaning
- Multi-Group Proportional Representations for Text-to-Image Models (CVPR 2025 · similarity 0.89)
- SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.83)
- Finetuning Text-to-Image Diffusion Models for Fairness (ICLR 2024 · similarity 0.82)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.81)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.81)
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.80)
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.