Algorithmic Fairness through Optimal Transport and Wasserstein Barycenters

This direction is merged into another direction: Wasserstein barycenter alignment. Its last version stays readable and citable here.

Version 1 of 164 papers100 conceptsfrozen 2026-10-04

In the quest to build equitable machine learning systems, researchers have converged on a powerful geometric intuition: if a model is biased, it is because the probability distributions of its predictions differ across demographic groups. To achieve "Statistical Parity"—where a model’s output is independent of a protected attribute like race or gender—we must find a way to align these disparate distributions.

Optimal Transport (OT) theory provides the mathematical machinery to do exactly this. By treating fairness as a problem of "moving" probability mass, researchers have developed a framework that is both theoretically rigorous and practically flexible.

The Geometry of Bias

Imagine the risk scores assigned by a lending algorithm. If we plot the distribution of these scores for two different groups, we might see two distinct "mountains" of probability density. If the mountain for Group A is shifted significantly toward higher risk than Group B, the model is statistically biased, even if it never explicitly "saw" the group labels.

As Eustasio del Barrio, Fabrice Gamboa, Jean-Michel Loubes, and Paula Gordaliza argue in Obtaining fairness using optimal transport theory, fairness can be achieved by transforming these biased input distributions into a shared reference distribution. This process, often called "data repair," ensures that protected attributes become indistinguishable within the feature space.

The Wasserstein Barycenter: A Fair Middle Ground

If we have multiple groups, where should we move their distributions? The most efficient target is the Wasserstein barycenter—the geometric "average" of all group distributions.

In Fair regression with Wasserstein barycenters, Evgenii Chzhen proves a fundamental result: the optimal fair predictor (the one that minimizes error while satisfying demographic parity) is mathematically equivalent to the Wasserstein-2 barycenter of the group-specific distributions.

This barycenter acts as a "fair center of mass." By mapping every group toward this central distribution, we minimize the total "transport cost"—the amount we have to distort the original data or predictions to achieve fairness. Aldo Pacchiano, Silvia Chiappa, and colleagues at DeepMind and Google further demonstrate in Wasserstein Fair Classification that using the barycenter as a target provides the most efficient path to parity, preserving as much of the original predictive information as possible.

Repairing without Retraining

One of the most significant advantages of the OT approach is that it can be applied as a post-processing step. This means we can "fix" a pre-trained, black-box model without needing to see its training data or change its internal weights.

Han Zhao and Ruicheng Xian have shown that this post-processing logic extends across tasks. In Fair and Optimal Classification via Post-Processing, they demonstrate that the most accurate fair classifier is found by identifying the geometric center of mass between group distributions within a probability simplex. Similarly, in Differentially Private Post-Processing for Fair Regression, they combine this with privacy constraints, using private histogram estimation to calculate the barycenter.

For practitioners, this means fairness is modular. As Qihang Lin and Tianbao Yang describe in Multi-Output Distributional Fairness via Post-Processing, a "Transporting to Approximate Barycenter" (TAB) module can be plugged into existing pipelines to align joint distributions in multi-task or multi-output settings.

Enforcing perfect fairness often comes at the cost of accuracy—a phenomenon known as the "price of fairness." OT theory allows us to navigate this trade-off with precision.

Instead of jumping all the way to the barycenter, we can move distributions along a geodesic—the shortest path in Wasserstein space. Silvia Chiappa and Ray Jiang in A General Approach to Fairness with Optimal Transport show that by choosing an intermediate point on this path, practitioners can "nudge" a model toward fairness while controlling the loss in utility. Paula Gordaliza formalizes this as "Random Repair," where the degree of alignment with the barycenter is a tunable parameter.

Beyond Simple Tables: Graphs and Privacy

The reach of OT-based fairness extends into complex data structures and privacy-sensitive environments:

Why It Works: The Intuition of Quantiles

At its heart, the OT approach to fairness works because it aligns quantiles. In a fair system, the person at the 90th percentile of Group A should receive the same outcome as the person at the 90th percentile of Group B.

As Jean-Michel Loubes and Philippe Rigollet explain in Projection to Fairness in Statistical Learning, mapping group-specific prediction quantiles to a shared barycenter ensures that relative merit is preserved while systemic group-level shifts are erased. This geometric projection is the most "conservative" way to achieve fairness, changing the original model only as much as is strictly necessary to satisfy the constraint of equality.

Papers in this version

Showing 60 of 64 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_algorithmic_fairness_through_optimal_tra_a71bb663,
  title = {Algorithmic Fairness through Optimal Transport and Wasserstein Barycenters},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/algorithmic-fairness-through-optimal-transport-and-wasserstein-barycenters/rdr_a71bb6633b4c44308e678b2d4f1db361},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_algorithmic_fairness_through_optimal_tra_a71bb663_v1,
  title = {Algorithmic Fairness through Optimal Transport and Wasserstein Barycenters},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/algorithmic-fairness-through-optimal-transport-and-wasserstein-barycenters/rdr_a71bb6633b4c44308e678b2d4f1db361/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}