Threshold Invariant Fairness for Machine Learning Models

43 concepts28 papers2 proposals

Most algorithmic fairness audits treat equity as a static property. A model outputs a continuous score—such as a credit risk score, a biometric similarity metric, or a recidivism probability—and a decision boundary is drawn at a default threshold (often 0.5). If the selection rates or error rates match across demographic groups at this specific boundary, the model is declared "fair."

However, as Mingliang Chen et al. argue in Towards Threshold Invariant Fair Classification, real-world decision-makers frequently adjust these thresholds due to changing economic conditions, resource constraints, or policy updates. If a model's fairness guarantees are tied to a single operational point, shifting the threshold even slightly can cause the model to become highly discriminatory. Matt Olfat et al. point out in Spectral Algorithms for Computing Fair Support Vector Machines that optimizing for a single threshold allows the underlying score distributions of protected groups to remain completely distinct, creating a hidden risk of post-training discrimination.

To resolve this, a growing body of research focuses on threshold-invariant fairness: designing models and metrics that guarantee equitable outcomes across all possible decision boundaries.

The Geometry of Threshold Sensitivity

To understand why traditional fairness metrics fail when thresholds shift, consider the probability density functions of risk scores assigned to two different demographic groups. Suppose Group A's scores form a narrow, tall bell curve centered at 0.5, while Group B's scores form a wide, flat bell curve also centered at 0.5.

If an algorithm sets the classification threshold exactly at 0.5, both groups will have an identical 50% acceptance rate, satisfying demographic parity perfectly. But what happens if a bank faces a capital shortage and raises the selection threshold to 0.7? Because Group B's distribution has fatter tails, a much larger proportion of Group B will qualify compared to Group A. Conversely, if the threshold drops to 0.3, Group A will suddenly see a massive surge in approvals relative to Group B.

This instability means that standard metrics like demographic parity difference ($\Delta DP$) are highly fragile. As Xiaotian Han et al. demonstrate in Retiring ΔDP: New Distribution-Level Metrics for Demographic Parity, evaluating fairness at a single point fails to capture these structural mismatches. To rigorously detect these discrepancies, Cyrus DiCiccio et al. introduce a permutation testing framework in Evaluating Fairness Using Permutation Tests, demonstrating that a model's statistically significant bias can fluctuate dramatically depending on the chosen operating point.

This threshold sensitivity presents severe operational risks across diverse domains:

This sensitivity even undermines compliance audits when protected attributes are unobserved. In Fairness Under Unawareness: Assessing Disparity When Protected Class Is Unobserved, Nathan Kallus et al. show that when proxy models are used to impute missing demographic data, small, arbitrary changes in the classification thresholds of those proxies lead to wildly inconsistent disparity estimates.

Achieving Invariance via Distributional Alignment

How do we build models that remain fair across all possible thresholds? The most direct solution is to force the continuous score distributions of different demographic groups to be mathematically identical. If the cumulative distribution functions (CDFs) of the scores match exactly across all groups, then drawing a line at any threshold $t$ will yield identical acceptance rates, achieving true threshold-invariant fairness.

One powerful framework for this is Optimal Transport. In Wasserstein Fair Classification, Aldo Pacchiano et al. introduce "Strong Demographic Parity." They use Wasserstein-1 distances to map each group's score distribution to a shared "group barycenter." Because the resulting distributions are perfectly aligned, the classifier remains fair no matter where an operational threshold is set post-training.

[Group A Scores] ---\
                     --> [Wasserstein Barycenter] --> [Perfectly Aligned Scores]
[Group B Scores] ---/       (Shared Distribution)       (Fair at any threshold)

A similar intuition drives post-processing techniques like "Geometric Repair." In Repairing Regressors for Fair Binary Classification at Any Decision Threshold, Kweku Kwegyir-Aggrey et al. adjust model scores prior to thresholding to maximize a new metric called Distributional Parity, guaranteeing consistent fairness across the entire spectrum of potential cutoffs.

Alternatively, this alignment can be performed during data pre-processing. Ivan Brugere et al. introduce a method in A Canonical Data Transformation for Achieving Inter- and Within-group Fairness that transforms the underlying data into a "canonical domain." This approach ensures that when a model is trained on the transformed data, it achieves threshold-invariant group fairness while strictly preserving the relative rankings of individuals within each group.

Bypassing Thresholds with Rank-Based Optimization

Instead of trying to align complex continuous distributions directly, another strategy bypasses thresholds entirely by optimizing for rank-based, threshold-agnostic objectives.

Consider the Area Under the ROC Curve (AUC), which measures a model's ability to correctly rank a randomly chosen positive instance above a randomly chosen negative instance. Because AUC evaluates a model across all possible thresholds simultaneously, optimizing for fair AUC ensures robust performance across all operational settings. Qihang Lin et al. leverage this in Stochastic Methods for AUC Optimization subject to AUC-based Fairness Constraints, formulating a scalable min-max optimization problem that enforces threshold-independent fairness. Similarly, Hortense Fong et al. show in Fairness for AUC via Feature Augmentation that strategically acquiring more predictive features for disadvantaged groups can directly bridge these holistic ranking gaps.

Ranks can also be used to construct distribution-free fairness guarantees. In FaiREE: fair classification with finite-sample and distribution-free guarantee, James Y. Zou et al. utilize the relative ranks of model scores within demographic subgroups rather than their raw values. Because ranks are uniformly distributed by definition, this approach allows practitioners to establish stable fairness thresholds without making any assumptions about the underlying data distribution.

This rank-based philosophy extends naturally to regression and information retrieval. Chen Zhao et al. demonstrate in Rank-Based Multi-task Learning for Fair Regression that non-parametric rank tests, such as the Mann-Whitney U statistic, can completely decouple continuous predictions from protected attributes. In a similar vein, Eugene Yang et al. introduce the PEER (Probability of Equal Expected Rank) metric in Language Fairness in Multilingual Information Retrieval, using non-parametric statistical tests to ensure that an item's expected rank is entirely independent of its language or group membership.

Key Foundational Work

Fairness Under Unawareness: Assessing Disparity When Protected Class Is Unobserved (2019) — contributed 2 concept elements to this direction.
Wasserstein Fair Classification (2019) — contributed 1 concept elements to this direction.
Spectral Algorithms for Computing Fair Support Vector Machines (2018) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_threshold_dependent_fairness_67d7172c,
  title = {threshold-dependent fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/threshold-dependent-fairness/rdr_67d7172c8891400a928ff195d3c8b33f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_threshold_dependent_fairness_67d7172c_v1,
  title = {threshold-dependent fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/threshold-dependent-fairness/rdr_67d7172c8891400a928ff195d3c8b33f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake