Distributional Fairness via Optimal Transport Theory

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

Version 1 of 124 papers26 conceptsfrozen 2026-10-04

Traditional algorithmic fairness often relies on "point-wise" metrics—checking if the average loan approval rate is the same for two groups or if a binary error rate is balanced. However, these averages can be deceptive. Two groups might have the same mean score but wildly different variances or distributions, leading to hidden inequities at the extremes.

To address this, a growing body of research leverages Optimal Transport (OT) theory to treat fairness as a geometric problem. Instead of matching simple averages, these methods aim to align the entire probability distributions of outcomes or data features.

Fairness as the Cost of Transformation

The core intuition behind this direction is that unfairness can be quantified as the "work" required to make two distributions identical. In OT, this is typically measured by the Wasserstein distance (also known as the Earth Mover’s Distance).

As Ann-Kristin Becker, Oana Dumitrasc, and Klaus Broelemann explain in Standardized Interpretable Fairness Measures for Continuous Risk Scores, the Wasserstein-1 distance provides a physical intuition: it represents the minimum effort needed to transform one group's score distribution into another's. This moves fairness from a binary "yes/no" property to a continuous, unit-based measure of disparity.

Fixing the Data, Not Just the Model

While many fairness interventions happen during model training, Eustasio del Barrio, Paula Gordaliza, and colleagues argue for "fixing the data" at the source. In Obtaining fairness using optimal transport theory, they propose using OT to "repair" biased datasets.

The mechanism works by finding the Wasserstein barycenter—a "middle ground" distribution that represents a fair version of the data. By mapping biased group distributions toward this barycenter, researchers can create a "Random Repair" method. This allows practitioners to precisely tune the trade-off between accuracy and fairness by deciding how far to move the data points toward the fair ideal.

Why Optimal Transport Works

Optimal transport offers several mathematical advantages over traditional divergence measures (like KL-divergence):

  1. Distributional Completeness: As noted by Luca Oneto and Silvia Chiappa, matching only the mean or variance fails to address inequities at different percentiles. OT ensures fairness across the entire range of outcomes.
  2. Monotonic Invariance: In Interpretable Distribution-Invariant Fairness Measures for Continuous Scores, Broelemann and Becker demonstrate that Wasserstein-based measures are invariant to monotonic transformations. This prevents "fairness washing," where a company might rescale scores to make them look fairer without actually changing the relative outcomes for individuals.
  3. Stable Optimization: In dynamic environments like MLOps, Yinghui Huang and Zihao Tang show in AdapFair that the Wasserstein distance provides a smoother, more stable optimization signal than discrete metrics, which is crucial for maintaining fairness as data drifts over time.

Auditing and Explaining Bias

Beyond enforcement, OT serves as a powerful tool for auditing. Nian Si, Jose Blanchet, and their team in Testing Group Fairness via Optimal Transport Projections frame auditing as a minimum-cost transportation problem. They calculate the smallest possible perturbation to a dataset required to satisfy fairness constraints; if that "distance to fairness" is large, the model is significantly biased.

This distributional approach even extends to complex data types. For example, Po-Sen Huang and Huan Zhang use Wasserstein-1 distance to measure sentiment bias in Large Language Models (Reducing Sentiment Bias in Language Models via Counterfactual Evaluation), while Yushun Dong and Jundong Li use it to attribute unfairness in Graph Neural Networks to specific training nodes (BIND).

By treating fairness as a problem of aligning probability "shapes" rather than just matching numbers, Optimal Transport provides a robust, geometrically grounded framework for building more equitable algorithms.

Papers in this version

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_distributional_fairness_via_optimal_tran_14c91e09,
  title = {Distributional Fairness via Optimal Transport Theory},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/distributional-fairness-via-optimal-transport-theory/rdr_14c91e09c78e4de2ab4034e6dd773416},
  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_distributional_fairness_via_optimal_tran_14c91e09_v1,
  title = {Distributional Fairness via Optimal Transport Theory},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/distributional-fairness-via-optimal-transport-theory/rdr_14c91e09c78e4de2ab4034e6dd773416/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}