Most fair machine learning models suffer from a hidden vulnerability: they are incredibly brittle to operational changes. When a risk-scoring model is deployed—whether for credit lending, hiring, or criminal justice—it typically outputs a continuous score or probability. To make a final binary decision (such as approving or denying a loan), a human decision-maker or business rule applies a classification threshold.
Standard fairness metrics, such as Demographic Parity, evaluate a model's outcomes at just a single, fixed threshold. However, as business constraints, economic climates, or risk tolerances shift, managers frequently adjust these thresholds. As highlighted by Chen et al., a model that appears perfectly equitable at a threshold of $0.5$ can exhibit severe demographic bias if the threshold is moved to $0.7$.
To resolve this, a growing body of research focuses on threshold-invariant fairness. The core insight is simple: to ensure a model remains fair across all possible deployment scenarios, we must shift our focus from balancing discrete decisions to aligning the underlying continuous score distributions.
The Intuition: Shifting from Points to Distributions
Why does changing a threshold destroy fairness? Consider a model that generates risk scores between $0$ and $1$ for two demographic groups. If the score distributions for these two groups have different shapes—even if they happen to intersect or balance out at one specific point—moving away from that point will immediately create a disparity in selection rates.
Standard Fairness (Single Threshold):
Group A Score Density: [---\___/---]
Group B Score Density: [---/~~~\---]
|
Fixed Threshold (Fair here, biased elsewhere)
Threshold-Invariant Fairness (Distributional Alignment):
Group A Score Density: [---/~~~\---]
Group B Score Density: [---/~~~\---]
| |
Any chosen threshold remains fair
To achieve threshold invariance, the model's continuous outputs must be statistically indistinguishable between protected groups. If the full probability density functions of the scores are identical across demographics, drawing a decision cutoff anywhere along the spectrum will naturally select an equal proportion of candidates from each group. As Kwegyir-Aggrey et al. note, this principle moves fairness from a superficial property of a tuned threshold to an intrinsic property of the scoring function itself.
Quantifying Fairness Across the Spectrum
To optimize models for threshold invariance, researchers have developed mathematical frameworks to measure the alignment between continuous distributions rather than discrete outcomes.
Cumulative and Area-Based Metrics
Instead of checking the demographic gap at a single cutoff, we can aggregate the gaps across all possible cutoffs. Han et al. proposed evaluating fairness by calculating the area between the cumulative density functions (CDFs) of different groups, introducing metrics like the Area Between Probability Curves (ABPC). Mathematically, integrating the difference in positive selection rates across every threshold from zero to one provides a robust, aggregate measure of bias that cannot be gamed by cherry-picking a specific cutoff.
Optimal Transport and Wasserstein Distance
Another prominent approach treats distributional alignment as a geometric problem. Pacchiano et al. and Kwegyir-Aggrey et al. leverage the Wasserstein distance (from optimal transport theory) to quantify the divergence between group score distributions. Minimizing the Wasserstein distance effectively "shapes" the probability distributions until they overlap perfectly, guaranteeing what the authors term Strong Demographic Parity.
ROC Curve Matching
When the goal is error-rate fairness (such as Equalized Odds) rather than outcome parity, threshold invariance requires aligning the groups' Receiver Operating Characteristic (ROC) curves. Vummintala et al. introduced $\epsilon$-Equalized ROC, a framework ensuring that the true positive and false positive rates for different groups remain close across the entire decision spectrum. Similarly, Olfat et al. framed fairness as making the model's score distributions structurally incapable of predicting the protected attribute, forcing the corresponding ROC curve to flatten into a diagonal line.
Mechanisms for Achieving Distributional Alignment
Enforcing threshold-invariant fairness can be integrated at different stages of the machine learning pipeline:
- Pre-processing: Brugere et al. designed a canonical data transformation that maps input features into a shared domain before training begins. This ensures inter-group distributional fairness while preserving the relative, within-group rankings of individuals.
- In-processing: Fairness constraints can be embedded directly into the model's loss function. For example, Olfat et al. developed spectral algorithms to bound the covariance of Support Vector Machine (SVM) predictions across groups during training. Meanwhile, Pacchiano et al. utilized an optimization penalty that pulls each group's score distribution toward a shared Wasserstein barycenter.
- Post-processing: When a model is already trained and cannot be altered, practitioners can apply post-hoc calibration. Kwegyir-Aggrey et al. demonstrate "Geometric Repair" methods that map biased continuous scores onto a fair, unified distribution. In a similar vein, the FROC framework by Vummintala et al. warps output scores post-hoc specifically to minimize the distance between group ROC curves.
By shifting the paradigm from single-threshold tuning to full score distribution alignment, these approaches ensure that machine learning models remain robustly equitable, no matter how real-world decision boundaries shift.
Key Foundational Work
Wasserstein Fair Classification (2019) — contributed 2 concept elements to this direction.
Spectral Algorithms for Computing Fair Support Vector Machines (2018) — contributed 2 concept elements to this direction.