Research map / Learning theory and algorithms / Fairness
Fairness without Harm: An Influence-Guided Active Sampling Approach
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 Fair Influential Sampling (FIS), an active data-sampling method that selects unlabeled examples by first-order influence on validation fairness and accuracy, using sensitive attributes only on a small validation set.
Problem
Improving group fairness (DP, EOp, EOd) of classifiers without harming accuracy, when acquiring more training data cannot require sensitive attribute annotations on the training or sampled data.
Approach
Score each unlabeled example by the first-order influence of adding it to training on validation-set accuracy loss and fairness loss (gradient inner products with the validation gradients), using proxy labels, so no training sensitive attributes are needed; keep only checkpoints whose validation accuracy stays within a tolerance of the warm-start model.
Key results
- On CelebA, FIS outperforms baselines on three fairness metrics while keeping accuracy at the same level as the ERM baseline (per the paper's Figure).
- On Adult, FIS achieves the lowest violation for equality of opportunity and comparable DP violation.
- On COMPAS, FIS gives a much better accuracy-fairness trade-off than other baselines, while JTT-20 reduces fairness violation at a significant accuracy cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FIS instead of the baselines | CelebA · CelebA | fairness violation (DP, EOp, and EOd) | Helped |
| Use FIS instead of the baselines | CelebA · CelebA | test accuracy | No clear effect |
| Use FIS instead of the baselines | Adult · Adult | EOp fairness violation | Helped |
| Use FIS instead of the baselines | Adult · Adult | DP fairness violation | No clear effect |
| Use FIS instead of the baselines | Compas · Compas | accuracy-fairness trade-off | Helped |
| Use JTT-20 instead of other baselines | Compas · Compas | fairness violation | Helped |
| Use JTT-20 instead of other baselines | Compas · Compas | test accuracy | Hurt |
| Vary the label budget (ablation) | FIS · CelebA | fairness violation | Helped |
| Vary the label budget (ablation) | FIS · CelebA | test accuracy | No clear effect |
| Use JTT-20 under a limited label budget (ablation) | JTT-20 · CelebA | fairness violation | Helped |
| Use JTT-20 under a limited label budget (ablation) | JTT-20 · CelebA | test accuracy: about 50% | Hurt |
| Vary the validation set size (ablation) | FIS · CelebA | test accuracy | No clear effect |
Surprising findings
- JTT-20 reaches near-zero fairness violation on CelebA under limited budget, but the paper argues its accuracy (about 50%) is uninformative.
Limitations
- Requires a clean validation set with sensitive attributes; sampling may be inefficient when the validation set is noisy.
- Influence is a first-order approximation, which is inaccurate for some examples.
- Proxy labels from the warm-start model may differ from true labels, requiring a while-loop re-selection.
Experimental setup
Datasets: CelebA UCI Adult COMPAS
Benchmarks: CelebA UCI Adult COMPAS
Metrics: test accuracy demographic parity violation equal opportunity violation equalized odds violation
Baselines: Base (ERM) Random BALD ISAL JTT-20
Builds on: influence functions active learning Just Train Twice ISAL
Closest papers by meaning
- Fairness via Independence: A General Regularization Framework for Machine Learning (ICLR 2026 · similarity 0.90)
- Active Sampling for Min-Max Fairness (ICML 2022 · similarity 0.90)
- Group Fairness by Probabilistic Modeling with Latent Fair Decisions (AAAI 2021 · similarity 0.89)
- Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.88)
- Achievable Fairness on Your Data With Utility Guarantees (NeurIPS 2024 · similarity 0.88)
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.88)
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.