Robust Fairness Across All Decision Thresholds

This direction is split into other directions: threshold-invariant fairness. Its last version stays readable and citable here.

Version 1 of 123 papers36 conceptsfrozen 2026-10-04

Most algorithmic fairness techniques are "brittle." A model might achieve perfect demographic parity at a standard classification threshold of 0.5, but as soon as a practitioner adjusts that threshold to account for different business costs or risk tolerances, the fairness guarantees evaporate. This instability occurs because standard methods only equalize the outcomes at a single point, ignoring the underlying distribution of scores the model assigns to different groups.

A growing body of research proposes a more robust alternative: threshold-agnostic fairness. By aligning the entire probability distribution of model outputs across protected groups, researchers ensure that a model remains fair regardless of where the final decision boundary is drawn.

The Problem with Single-Point Fairness

In a typical classification task, a model produces a continuous score (e.g., a probability between 0 and 1), which is then converted into a binary decision by a threshold. Standard fairness metrics like "Equalized Odds" or "Demographic Parity" are usually calculated based on these final binary decisions.

However, as Aldo Pacchiano, Silvia Chiappa, and their colleagues argue in Wasserstein Fair Classification, this "point-wise" approach is insufficient. If the score distributions for two groups have different shapes—even if they result in the same percentage of positive outcomes at $p=0.5$—shifting the threshold to $p=0.7$ might suddenly favor one group over the other. To solve this, they introduce Strong Demographic Parity, which requires the full cumulative distribution functions (CDFs) of the scores to be identical across groups.

Aligning Distributions via Optimal Transport

The most direct way to achieve threshold-agnostic fairness is to force the model's output distributions to match. Aldo Pacchiano and Silvia Chiappa utilize Wasserstein distances to measure the "work" required to transform one group's score distribution into another's. By minimizing this distance to a shared "barycenter" (a middle-ground distribution), they create a model where the scores are statistically indistinguishable between groups.

This concept is extended to decentralized settings in WassFFed, where Zhongxuan Han and Chaochao Chen use Wasserstein barycenters to align model outputs in Federated Learning. This prevents bias even when data is distributed unevenly across different clients. Similarly, Ivan Brugere and colleagues propose a Canonical Data Transformation that maps data into a shared domain where distributions are aligned, preserving the relative ranking of individuals while satisfying group-level fairness.

Fairness as Indistinguishable Ranking

Another way to think about threshold-agnostic fairness is through the lens of ranking. If a model is truly fair across all thresholds, its scores should provide zero predictive power regarding an individual's protected attribute.

Matt Olfat and Anil Aswani frame this in Spectral Algorithms for Computing Fair Support Vector Machines. They suggest that a perfectly fair classifier should result in a Receiver Operating Characteristic (ROC) curve for the protected attribute that follows the diagonal line. If an adversary cannot use your model's scores to guess whether a person is from Group A or Group B better than random chance, the model is fair at every possible cutoff.

This ranking-based intuition leads naturally to optimizing the Area Under the ROC Curve (AUC). Qihang Lin and Tianbao Yang have developed stochastic methods to maximize a model's performance while enforcing AUC-based fairness constraints. Because AUC is inherently threshold-independent, these constraints ensure equitable performance across all possible decision boundaries.

Distribution-Free and Non-Parametric Robustness

To make these ideas practical for real-world data, which is often messy or heavy-tailed, researchers have turned to non-parametric statistics.

  • Rank-Based Tests: Chen Zhao and Feng Chen use the Mann-Whitney U statistic in Rank-Based Multi-task Learning to eliminate dependencies between predictions and protected attributes. Because it relies on ranks rather than raw values, it is robust to skewed data.
  • Order Statistics: In FaiREE, Linjun Zhang and James Zou leverage order statistics to provide fairness guarantees that are "distribution-free." This ensures that fairness holds even with small sample sizes and unknown data distributions.
  • Margin Alignment: When models are allowed to "abstain" from difficult predictions, they often become less fair. Joshua K. Lee and colleagues show in Fair Selective Classification Via Sufficiency that aligning the margin distributions (the confidence scores) across groups ensures that a single rejection threshold treats all groups equitably.

Stability in Deployment

The shift toward distribution-level fairness isn't just a theoretical preference; it has major implications for model deployment. Alan Mishler and Niccolo Dalmasso demonstrate that standard fairness measures are unstable in "performative" settings where model predictions change future data. They find that Demographic Parity is uniquely stable because it focuses on the distribution of decisions relative to attributes, ignoring the feedback loops that can make other metrics fluctuate after deployment.

By moving from point-wise constraints to full distribution alignment, these researchers are creating models that don't just look fair on a static test set, but remain fair in the fluid, unpredictable environments of the real world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_robust_fairness_across_all_decision_thre_af1c7d59,
  title = {Robust Fairness Across All Decision Thresholds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-fairness-across-all-decision-thresholds/rdr_af1c7d59b5cf48dda8cadd46a3420d6c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_fairness_across_all_decision_thre_af1c7d59_v1,
  title = {Robust Fairness Across All Decision Thresholds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-fairness-across-all-decision-thresholds/rdr_af1c7d59b5cf48dda8cadd46a3420d6c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}