Research map / Learning theory and algorithms / Fairness
Fairness under Covariate Shift: Improving Fairness-Accuracy Tradeoff with Few Unlabeled Test Samples
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 an unsupervised test-adaptation method for fair classification under covariate shift, combining a weighted-entropy objective on a few unlabeled test samples (with implicit importance-ratio estimation) and Wasserstein representation matching, and introduces asymmetric covariate shift.
Problem
Fair classification under covariate shift when only a small set of unlabeled test samples is available, where existing importance-weighting and density-ratio methods have high variance and fairness-accuracy tradeoffs degrade.
Approach
A min-max weighted entropy objective on unlabeled test samples, with weights constrained to implicitly match importance ratios (no density estimation), combined with training ERM and a Wasserstein-2 representation matching fairness regularizer; the bound's generalization does not depend on importance-weight variance.
Key results
- On Adult, the method reaches about 15% error with equalized odds about 0.075, where the closest baselines MLP and RF fall short on one of the metrics.
- Under asymmetric shift on Drug, the method is reported as about 10x and 5x better than the two importance-sampling baselines on equalized odds.
- Adding the weighted entropy term reduces error on Drug from 35.9% to 31.9% and equalized odds from 0.076 to 0.006 in the ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the weighted entropy formulation (ablation) | Proposed method with Wasserstein objective · Arrhythmia, asymmetric shift A=0 | Error %: 28.648 (3.079) -> 27.648 (2.978) | Helped |
| add the weighted entropy formulation (ablation) | Proposed method with Wasserstein objective · Drug, asymmetric shift A=0 | Equalized Odds: 0.076 (0.043) -> 0.006 (0.013) | Helped |
| increase entropy regularizer λ1 from 0 to 0.1 (ablation) | Proposed method · Adult | Error (in %): 23.819 (8.593) -> 14.626 (1.318) | Helped |
| increase entropy regularizer λ1 from 0.1 to 1.0 (ablation) | Proposed method · Adult | Error (in %): 14.626 (1.318) -> 14.787 (1.326) | No clear effect |
| increase Wasserstein regularizer λ2 to 1.0 (ablation) | Proposed method · Adult | Error (in %): 14.787 (1.326) -> 42.280 (32.581) | Hurt |
| replace weighted entropy with unweighted entropy (ablation) | Proposed method without Wasserstein objective · Communities and Crime | Error %: 34.291 (4.463) -> 45.787 (12.900) | Hurt |
| replace weighted entropy with unweighted entropy (ablation) | Proposed method with Wasserstein objective · Communities and Crime | Equalized Odds: 0.328 (0.073) -> 0.201 (0.200) | Helped |
Surprising findings
- Unweighted entropy minimization attains better equalized odds but only by producing a nearly random classifier (about 50% error on Communities).
- Naively combining a state-of-the-art distributionally robust method (RGD-Exp) with Wasserstein representation matching does not achieve a good fairness-accuracy tradeoff.
Limitations
- Bound assumes the predictor approximates the true soft predictor within a constant factor epsilon, which is enforced heuristically by pre-training.
- Experiments are limited to small tabular datasets with only 50 unlabeled test samples.
- Results depend on several hyperparameters (lambda1, lambda2, epsilon) tuned by grid search.
Experimental setup
Datasets: Adult Arrhythmia Communities and Crime Drug
Benchmarks: Adult Arrhythmia Communities and Crime Drug
Metrics: error rate equalized odds accuracy parity
Baselines: MLP Adversarial Debiasing Robust Fair Robust Shift Fair Z-Score Adaptation KLIEP LSIF RGD-Exp
Builds on: adversarial debiasing importance weighting KLIEP LSIF RGD-Exp
Closest papers by meaning
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.86)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.86)
- Mitigating Spurious Correlations via Disagreement Probability (NeurIPS 2024 · similarity 0.86)
- Size-adaptive Hypothesis Testing for Fairness (NeurIPS 2025 · similarity 0.86)
- A Sequentially Fair Mechanism for Multiple Sensitive Attributes (AAAI 2024 · similarity 0.86)
- On the Burden of Achieving Fairness in Conformal Prediction (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.