Research map / Learning theory and algorithms / Optimal transport
Auditing and Enforcing Conditional Fairness via Optimal Transport
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 conditional demographic disparity measures based on optimal transport (Wasserstein and l_p aggregation) and two regularizers, FairBiT (bi-causal transport) and FairLeap, to enforce conditional demographic parity for many-level and continuous-output settings.
Problem
Auditing and enforcing conditional demographic parity (CDP) when the legitimate feature has many levels and/or model outputs are continuous; existing methods only match first moments or use proxies that do not bound conditional disparity.
Approach
Define conditional demographic disparity via nested Wasserstein distances (equivalent to bi-causal transport distance) and an l_p aggregation of level-wise Wasserstein distances, then use differentiable nested Sinkhorn divergences as regularizers during training so that full conditional output distributions are matched.
Key results
- FairBiT has points on the Pareto frontier for every experiment when CDD is measured by the Wasserstein-sense disparity, generally outperforming other methods.
- FairLeap (uniform) lies on the Pareto frontier for all 4 datasets when CDD is measured in the l_p sense with uniform Q and p=1.
- Preprocessing repair attains predictive performance on par with no regularization but fails to enforce conditional fairness.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FairBiT regularization | MLP · Drug, Adult, Law School, Communities and Crime | fairness-predictive power trade-off | Helped |
| Use FairLeap (uniform) regularization | MLP · Drug, Adult, Law School, Communities and Crime | fairness-predictive power trade-off | Helped |
| Use FairLeap variants with alternative aggregation weights | MLP · Drug, Adult, Law School, Communities and Crime | fairness-predictive power trade-off | Helped |
| Modify adversarial debiasing to pass (f(X), L) to the adversary | MLP · Adult, Drug, Law School | predictive performance and conditional fairness | Hurt |
| Modify adversarial debiasing to pass (f(X), L) to the adversary | MLP · Communities and Crime | conditional demographic disparity | Helped |
| Apply preprocessing repair | MLP · Drug, Adult, Law School, Communities and Crime | conditional fairness | Hurt |
| Use DCFR | MLP · Law School | CDD in the ℓp sense | Helped |
| Use DCFR | MLP · Drug, Adult, Communities and Crime | fairness-predictive power trade-off | Hurt |
| Use Wasserstein regularization | MLP · Law School, Communities and Crime | predictive power at matched CDD | Hurt |
| Use Wasserstein regularization | MLP · Drug, Adult | fairness-predictive power trade-off | Mixed |
| Train using only the legitimate feature L | MLP · Drug, Adult, Law School, Communities and Crime | predictive performance | Hurt |
Surprising findings
- DCFR, the existing state of the art for CDP, minimizes a proxy quantity whose small values do not imply small conditional disparity.
- Methods targeting demographic parity do not necessarily achieve conditional demographic parity.
Limitations
- Requires L to be discrete or discretized and multiple samples per level; continuous legitimate features need binning.
- Adversarial debiasing and DCFR comparisons rely on modified versions; sensitive feature access is needed for some baselines.
- Computational cost of nested Sinkhorn is quadratic in sample size.
Experimental setup
Datasets: Drug Adult Law School Communities and Crime
Benchmarks: Drug Adult Law School Communities and Crime
Metrics: AUC MSE conditional demographic disparity (Wasserstein sense) CDD (l_p sense)
Baselines: DCFR Adversarial debiasing Pre-processing repair Wasserstein regularization No regularization Legitimate only
Builds on: DCFR adversarial debiasing Sinkhorn divergence bi-causal transport distance
Closest papers by meaning
- Meta Optimality for Demographic Parity Constrained Regression via Post-Processing (ICML 2025 · similarity 0.87)
- Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.87)
- Demographic parity in regression and classification within the unawareness framework (NeurIPS 2026 · similarity 0.86)
- Regression under demographic parity constraints via unlabeled post-processing (NeurIPS 2024 · similarity 0.85)
- Size-adaptive Hypothesis Testing for Fairness (NeurIPS 2025 · similarity 0.84)
- On the Maximal Local Disparity of Fairness-Aware Classifiers (ICML 2024 · similarity 0.84)
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.