Research map / Label-efficient and robust learning / Recognition and long-tail
Navigating Towards Fairness with Data Selection
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 a fair data selection method that extends RHO-LOSS to fairness by using a CLIP zero-shot predictor as a proxy holdout model, corrected with a peer-prediction loss, to select training data less affected by label bias.
Problem
Fairness under label bias: existing fair-learning methods modify models or the training process and lack flexibility on large-scale datasets, while data selection methods need a clean holdout set.
Approach
Replace the RHO-LOSS holdout model with a zero-shot pre-trained predictor, make its loss fair via a peer-prediction term with cross-group pairing, and add resampling to correct selection bias, yielding a model-agnostic online batch selection rule that needs no holdout set or noise-rate estimate.
Key results
- On CelebA with 40% symmetric label bias, the method reaches 85.2% accuracy with DP violation 0.20, versus 79.7% and 0.28 for RHO-LOSS.
- Reaches 70% test accuracy on CelebA(0.4) in 29 epochs versus 46 for RHO-LOSS and 94 for uniform sampling.
- Ablations show robustness to zero-shot predictor (CLIP-RN50, ViT-B/16, RHO-LOSS validation model) and backbone (ResNet-18/50, DenseNet-121).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use fair data selection with peer prediction and resampling instead of RHO-LOSS | ResNet-18 classifier · LFW+a (0.2 label bias) | accuracy (%): 89.5$_{\pm 0.5}$ -> 90.9$_{\pm 0.6}$ | Helped |
| use fair data selection with peer prediction and resampling instead of RHO-LOSS | ResNet-18 classifier · LFW+a (0.4 label bias) | accuracy (%): 83.9$_{\pm 1.8}$ -> 88.7$_{\pm 0.7}$ | Helped |
| use fair data selection with peer prediction and resampling instead of Uniform | ResNet-50 classifier · CelebA (0.2 label bias) | accuracy (%): 82.0$_{\pm 1.3}$ -> 86.5$_{\pm 0.6}$ | Helped |
| use fair data selection with peer prediction and resampling instead of Uniform | ResNet-50 classifier · CelebA (0.4 label bias) | accuracy (%): 81.0$_{\pm 3.1}$ -> 85.2$_{\pm 1.7}$ | Helped |
| add batch resampling (ablation) | Ours · LFW+a and CelebA, both label-bias settings | accuracy (%): 90.0$_{\pm 1.3}$ -> 90.9$_{\pm 0.6}$; 86.2$_{\pm 0.7}$ -> 88.7$_{\pm 0.7}$; 85.5$_{\pm 0.6}$ -> 86.5$_{\pm 0.6}$; 84.5$_ | Helped |
| replace CLIP-RN50 with ViT-B/16 as zero-shot predictor (ablation) | Ours · CelebA (0.2 label bias) | accuracy (%): 85.5$_{\pm 0.6}$ -> 86.4$_{\pm 0.2}$ | Helped |
| replace CLIP-RN50 with ViT-B/16 as zero-shot predictor (ablation) | Ours · CelebA (0.4 label bias) | accuracy (%): 85.2$_{\pm 1.7}$ -> 84.5$_{\pm 1.2}$ | Hurt |
| replace ResNet-50 with ResNet-18 backbone (ablation) | Ours · CelebA | accuracy (%): 86.5$_{\pm 0.6}$ -> 84.8$_{\pm 1.4}$; 85.2$_{\pm 1.7}$ -> 84.1$_{\pm 0.2}$ | Hurt |
| vary γ upward (ablation) | Ours · CelebA (0.4 label bias) | accuracy | Helped |
| vary γ upward (ablation) | Ours · CelebA (0.4 label bias) | fairness violation | Helped |
| vary α (ablation) | Ours · CelebA (0.4 label bias) | fairness violation | No clear effect |
| increase selection ratio from 0.1 to 0.3 (ablation) | Ours · CelebA (0.4 label bias) | accuracy | Helped |
Surprising findings
- Gradient-norm selection performs worse than uniform sampling, especially under high label bias.
Limitations
- Fair distribution assumption is idealized and not observed in practice.
- Evaluated only on two image datasets with symmetric synthetic label bias.
- Gains over RHO-LOSS are small in several settings and some results have large variance.
Experimental setup
Datasets: CelebA LFW+a
Benchmarks: CelebA LFW+a
Metrics: accuracy demographic parity violation equal opportunity difference p%-rule epochs to target accuracy
Baselines: Uniform Grad Norm Grad Norm IS RHO-LOSS
Builds on: RHO-LOSS peer prediction Query By Committee
Closest papers by meaning
- Mitigating Label Bias in Machine Learning: Fairness through Confident Learning (AAAI 2024 · similarity 0.89)
- Fairness via Independence: A General Regularization Framework for Machine Learning (ICLR 2026 · similarity 0.88)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.87)
- Understanding Fairness and Prediction Error through Subspace Decomposition and Influence Analysis (NeurIPS 2025 · similarity 0.86)
- Utility-Fairness Trade-Offs and How to Find Them (CVPR 2024 · similarity 0.85)
- How Far Can Fairness Constraints Help Recover From Biased Data? (ICML 2024 · similarity 0.85)
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.