Research map / Learning theory and algorithms / Fairness
A Sequentially Fair Mechanism for Multiple Sensitive Attributes
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 sequential post-processing framework based on multi-marginal 2-Wasserstein barycenters that achieves demographic parity (exact and approximate) across multiple sensitive attributes, with a closed-form optimal fair predictor and a data-driven estimator.
Problem
Achieving demographic parity fairness for regression and classification when multiple sensitive attributes are present, where single-attribute fairness methods fail to address intersectional bias and combined-attribute encodings lack prioritization and interpretability.
Approach
Extends the single-attribute Wasserstein-barycenter closed-form fair predictor to multiple sensitive attributes by exploiting univariate barycenter associativity, so fairness can be enforced sequentially in any order (and with epsilon-relative-improvement trade-offs per attribute) with exact order-invariance of the final result.
Key results
- On folktables income regression, post-processing reduces unfairness from 0.378 to 0.019 at a cost of MSE rising from 0.547 to 0.596.
- With two sensitive attributes on folktables income classification, the method reaches unfairness 0.009 with F1 0.73, versus Fairlearn at unfairness 0.207 with F1 0.73, while running in about 10.7 s versus 355 s.
- On the Bank benchmark, the method achieves unfairness 0.03 at F1 0.90, compared with FairBalance at F1 0.94 and unfairness 0.16.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply sequential fairness post-processing | LightGBM · folktables Income regression | Unfairness: 0.378 $\pm$ 0.007 -> 0.019 $\pm$ 0.005 | Helped |
| Apply sequential fairness post-processing | LightGBM · folktables Income regression | MSE: 0.547 $\pm$ 0.003 -> 0.596 $\pm$ 0.003 | Hurt |
| Apply sequential fairness post-processing | LightGBM · folktables Income classification, one sensitive attribute | Unfairness: 0.170$\pm$ 0.001 -> 0.003 $\pm$ 0.002 | Helped |
| Apply sequential fairness post-processing | LightGBM · folktables Income classification, one sensitive attribute | F1: 0.753$\pm$ 0.001 -> 0.737 $\pm$ 0.001 | Hurt |
| Apply sequential fairness post-processing | LightGBM · folktables Income classification, two sensitive attributes | Unfairness: 0.354 $\pm$ 0.006 -> 0.009 $\pm$ 0.005 | Helped |
| Apply sequential fairness post-processing | LightGBM · folktables Income classification, two sensitive attributes | F1: 0.753$\pm$ 0.001 -> 0.73 $\pm$ 0.001 | Hurt |
| Apply sequential fairness post-processing | LightGBM · folktables Coverage classification | Unfairness: 0.127 $\pm$ 0.0 -> 0.011 $\pm$ 0.0 | Helped |
| Apply sequential fairness post-processing | LightGBM · folktables Coverage classification | F1: 0.587 $\pm$ 0.0 -> 0.584 $\pm$ 0.0 | Hurt |
| Use the sequential fairness method instead of Fairlearn | folktables Income classification, two sensitive attributes · folktables Income classification, two sensitive attributes | Unfairness: 0.207 $\pm$ 0.005 -> 0.009 $\pm$ 0.005 | Helped |
| Use the sequential fairness method instead of Fairlearn | folktables Coverage classification · folktables Coverage classification | Unfairness: 0.119 $\pm$ 0.0 -> 0.011 $\pm$ 0.0 | Helped |
Surprising findings
- Intermediate sequential fairness steps can increase unfairness on other sensitive attributes (fairness gerrymandering), even though the final exact-fair solution is identical regardless of correction order.
Limitations
- Analysis restricted to demographic parity; extension to label-conditional metrics such as equalized odds left open.
- Post-processing reduces predictive performance (F1 and MSE drop), and intermediate fairness steps incur performance loss.
- Requires discrete sensitive attributes and assumes non-atomic conditional distributions with finite second moments.
Experimental setup
Datasets: folktables (US Census, sunbelt states) Bank Adult tunadromd synthetic Gaussian data
Benchmarks: folktables Income (regression and classification) folktables Coverage Bank Adult tunadromd
Metrics: MSE F1-score demographic parity unfairness compute time
Baselines: Uncorrected model Fairlearn ExponentiatedGradient FairBalance
Builds on: Wasserstein barycenter fair predictor (single sensitive attribute) Fairlearn ExponentiatedGradient LightGBM
Closest papers by meaning
- The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine Learning (NeurIPS 2024 · similarity 0.89)
- Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.89)
- FairICP: Encouraging Equalized Odds via Inverse Conditional Permutation (ICML 2025 · similarity 0.89)
- Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.88)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.88)
- Post-hoc bias scoring is optimal for fair classification (ICLR 2024 · similarity 0.87)
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.