Research map / Learning theory and algorithms / Fairness
FairWASP: Fast and Optimal Fair Wasserstein Pre-processing
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 FairWASP, a pre-processing method that learns integer sample weights minimizing Wasserstein distance to the original data subject to empirical demographic parity, solved efficiently via LP dual and cutting plane methods.
Problem
Reducing demographic disparity in classification datasets via pre-processing without modifying the original features or labels, while keeping the reweighted data close to the original distribution.
Approach
Reformulate reweighting as a Wasserstein-minimizing mixed-integer program under demographic parity constraints; show the LP relaxation's dual is a low-dimensional convex problem solvable with cutting planes, and that the LP solution is integral so integer weights (duplicating/deleting samples) are optimal.
Key results
- FairWASP solves the reformulated problem within 5 seconds for all synthetic sizes up to n=12,800, while Gurobi and Mosek exceed the 1-hour limit when n > 10,000.
- FairWASP and FairWASP-PW are consistently on the Pareto frontier of the fairness-utility tradeoff across Adult, Communities and Crime, and Drug datasets, significantly improving over the uniform baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FairWASP instead of training on the original dataset as-is | MLP classifier · Adult, Communities and Crime, Drug, and synthetic datasets | fairness-utility tradeoff | Helped |
| Use the FairWASP optimization algorithm instead of commercial solvers | Synthetic dataset optimization · Synthetic dataset | runtime: within 5 seconds; commercial solvers exceed 1 hour when n > 10,000 | Helped |
| Use FairWASP instead of Gurobi or Mosek | Synthetic dataset optimization · Synthetic dataset, n > 10,000 | runtime: 1 hour | Helped |
Limitations
- Fairness guarantees are for empirical demographic parity only; finite-sample properties of downstream fairness-utility tradeoff not characterized.
- FairWASP solutions are less accurate than solver solutions due to looser tolerances (relative duality gap 1e-3).
- Optimized preprocessing only applicable to small binarized datasets, so compared only on Adult.
Experimental setup
Datasets: Adult Drug Communities and Crime German Credit synthetic dataset
Benchmarks: Adult Drug Communities and Crime German Credit synthetic dataset
Metrics: runtime AUC demographic disparity relative objective gap fairness violation
Baselines: Gurobi Mosek Disparate Impact Remover Learning Fair Representations Reweighing Optimized preprocessing Uniform
Builds on: Kamiran and Calders reweighing Wasserstein distance / optimal transport analytic center cutting plane method
Closest papers by meaning
- Fair Wasserstein Coresets (NeurIPS 2024 · similarity 0.89)
- Who’s the (Multi-)Fairest of Them All: Rethinking Interpolation-Based Data Augmentation Through the Lens of Multicalibration (AAAI 2025 · similarity 0.87)
- Post-hoc bias scoring is optimal for fair classification (ICLR 2024 · similarity 0.87)
- Efficient Fairness-Performance Pareto Front Computation (NeurIPS 2025 · similarity 0.86)
- FairDD: Fair Dataset Distillation (NeurIPS 2025 · similarity 0.86)
- A Sequentially Fair Mechanism for Multiple Sensitive Attributes (AAAI 2024 · 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.