Distributional Fairness through Optimal Transport and Moment Matching

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

Version 1 of 134 papers42 conceptsfrozen 2026-10-04

Most algorithmic fairness interventions rely on a simple intuition: if we want to be fair, we should make sure the "average" outcome is the same for everyone. If a bank’s loan model approves 20% of men and 20% of women, it might appear fair on the surface.

However, as researchers like Aldo Pacchiano and Silvia Chiappa argue in Wasserstein Fair Classification, averages are a dangerous abstraction. Two groups can share the same mean score while having radically different outcome distributions. One group might have scores tightly clustered around the average, while another faces high volatility or extreme "tails." This research direction moves beyond simple mean-matching toward Distributional Fairness—ensuring that the entire shape of the model's output is consistent across demographic groups.

The Failure of First Moments

In statistics, the mean is the "first moment." Many early fairness constraints focused exclusively on this moment. But as Jun Sakuma and colleagues point out in Recommendation Independence, matching only the first moment fails to account for variance (the second moment).

If a recommendation system provides the same average rating to different groups but the variance for one group is much higher, that group experiences a less predictable and potentially riskier service. In dimensionality reduction, Junghyun Lee demonstrates in Fair Streaming Principal Component Analysis that failing to align these second-moment or covariance structures allows sensitive information to remain encoded in the data's spread, even if the centers are aligned.

The problem scales with complexity. Matt Olfat and Anil Aswani show in Spectral Algorithms for Computing Fair Support Vector Machines that a model might look fair at one specific decision threshold but exhibit massive disparities if the threshold is shifted. To be truly fair, a model must be fair across all possible thresholds.

Optimal Transport as a Fairness Nudge

If matching moments is like checking a few points on a map, Optimal Transport (OT) is like aligning the entire landscape. OT provides a mathematical toolkit to measure the distance between two probability distributions—most notably the Wasserstein distance.

In A General Approach to Fairness with Optimal Transport, Silvia Chiappa and her co-authors frame fairness as a distribution-matching problem. Instead of forcing a single number to match, they use the Wasserstein distance to "nudge" the entire prediction distribution of a protected group toward a target "fair" distribution (often the barycenter of all groups).

This approach works because the Wasserstein distance captures the geometry of the data. It doesn't just ask "how different are these averages?" but "how much work would it take to transform one distribution into the other?" This sensitivity allows researchers like Eustasio del Barrio to develop rigorous statistical tests for fairness, as seen in A Central Limit Theorem for $L_p$ transportation cost with applications to Fairness Assessment in Machine Learning.

Protecting the Tails and Local Disparities

A critical motivation for distributional fairness is the "glass ceiling" effect. Meichen Liu and colleagues observe in Conformalized Fairness via Quantile Regression that bias is often more pronounced at the extremes of a distribution—such as high-income brackets or top-tier performance levels. By using quantile regression and OT, they ensure that fairness holds not just for the median individual, but across every percentile.

Similarly, Jinqiu Jin and Fuli Feng argue in On the Maximal Local Disparity of Fairness-Aware Classifiers that average differences can mask "worst-case" local disparities. Their MCDP metric identifies specific regions of the prediction space where groups are treated most differently, forcing the model to be robustly fair everywhere, not just on average.

Beyond Wasserstein: Kernel Methods and MMD

While Optimal Transport is a dominant theme, it is not the only way to achieve distributional parity. Another powerful tool is the Maximum Mean Discrepancy (MMD). MMD measures the distance between distributions by mapping them into a high-dimensional feature space (a Reproducing Kernel Hilbert Space).

Junghyun Lee and Chang D. Yoo utilize this in Fast and Efficient MMD-Based Fair PCA. By minimizing MMD, they can penalize higher-order distributional differences that simple covariance constraints might miss. This is particularly useful in generative tasks; for instance, Aditya Varun V uses an MMD regularizer in Learning Counterfactually Fair Models via Improved Generation with Neural Causal Models to ensure that "what-if" scenarios are distributionally equal across groups.

Why It Works: The Intuition of Alignment

The shift from mean-matching to distributional alignment works because it treats fairness as a structural property of the model rather than a post-hoc correction.

  1. Invariance to Scaling: As Ann-Kristin Becker notes in Standardized Interpretable Fairness Measures for Continuous Risk Scores, Wasserstein-based measures are invariant to monotonic transformations. This prevents "fairness washing," where a developer might mathematically rescale scores to make means look closer without actually changing the relative treatment of individuals.
  2. Preserving Diversity: In image tasks, Guy Ohayon shows in Perceptual Fairness in Image Restoration that matching distributions ensures an algorithm doesn't "collapse" the variety of outputs for a specific demographic, maintaining quality and diversity for all.
  3. Risk Parity: By aligning the full distribution, models ensure that different groups face similar levels of uncertainty and risk, a concept explored by Grani A. Hanasusanto in Distributionally Fair Stochastic Optimization using Wasserstein Distance.

Ultimately, this research direction suggests that for an algorithm to be truly fair, it must not only get the "average" right but must also ensure that the entire spectrum of opportunity and risk is distributed equitably.

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_through_optimal__61d6d26b,
  title = {Distributional Fairness through Optimal Transport and Moment Matching},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/distributional-fairness-through-optimal-transport-and-moment-matching/rdr_61d6d26b32a048ba9c18becddc6d91ee},
  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_through_optimal__61d6d26b_v1,
  title = {Distributional Fairness through Optimal Transport and Moment Matching},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/distributional-fairness-through-optimal-transport-and-moment-matching/rdr_61d6d26b32a048ba9c18becddc6d91ee/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}