Algorithmic Fairness for Continuous Outcomes

28 concepts22 papers1 proposal

Most foundational work in algorithmic fairness focuses on binary classification—deciding whether an applicant receives a loan, an interview, or a bail hearing. In these discrete settings, fairness is a matter of counting and balancing rates across groups. However, many high-stakes real-world decisions rely on regression models that output continuous scores, such as credit scores, insurance premiums, risk assessments, or salary predictions (Komiyama et al.; Yeom et al.).

Enforcing fairness in regression is fundamentally harder than in classification. Because continuous variables have infinite possible values, traditional discrete fairness metrics do not naturally scale (Agarwal et al.; Chi et al.). Over the past several years, researchers have developed specialized mathematical frameworks to bridge this gap, moving from simple statistical averages to distribution-level alignment.

Why Simple Averages Fail

A naive approach to continuous fairness might simply equalize the expected value (the mean prediction) across different demographic groups. For example, ensuring that the average predicted risk score for Group A equals the average predicted risk score for Group B.

As Ghassemi et al. point out, this mean-matching strategy fails to capture the full extent of disparate impact. Two groups can share the exact same mean prediction while having completely different distribution shapes, variances, or tail behaviors.

If a model's predictions for Group A are tightly clustered around the mean, but its predictions for Group B are highly volatile and spread out, individuals in Group B will be disproportionately represented at the extreme ends of the scoring spectrum. In a lending scenario, this variance mismatch means that the highest-scoring individuals (who receive the best interest rates) and the lowest-scoring individuals (who are denied entirely) will be heavily skewed by demographic group, despite the model being "fair" on average.

To achieve true demographic parity, a system must align the entire shape of the continuous output distributions across groups.

Distribution Alignment via Optimal Transport

To enforce demographic parity across entire continuous distributions, researchers have turned to optimal transport theory. The goal is to find a way to transform the group-specific, biased prediction distributions into a single, shared, fair distribution while altering the original predictions as little as possible.

Chzhen et al. proved that the mathematically optimal fair predictor under demographic parity constraints is equivalent to the Wasserstein barycenter of the distributions from the different sensitive groups.

Imagine the prediction distributions of Group A and Group B as two different piles of earth. The Wasserstein barycenter is the "average" pile that minimizes the physical effort (transport cost) required to shovel both piles into a single, identical shape.

Group A Distribution  \
                       --->  [ Wasserstein Barycenter ]  ---> Fair Predictions
Group B Distribution  /         (Optimal Compromise)

By framing fair regression as a post-processing problem over Wasserstein spaces, the authors showed that you can take any standard, off-the-shelf regression model and map its biased continuous outputs to this barycenter. This guarantees that the final output distributions are identical across groups, minimizing the loss in predictive accuracy without requiring retraining.

Building on this geometric foundation, Xian et al. extended the framework to handle privacy constraints. They utilized private histogram estimation to compute these Wasserstein barycenters, proving that models can be made both distributionally fair and differentially private without exposing sensitive demographic data during the post-processing phase.

Conditioning on Continuous Targets

While Demographic Parity aims to make predictions independent of a sensitive attribute ($ \hat{Y} \perp A $), another core fairness criterion is Equalized Odds, which requires predictions to be independent of the sensitive attribute conditional on the true outcome ($ \hat{Y} \perp A \mid Y $). Equalized Odds ensures that the model's error rates and accuracy profiles are distributed equitably.

In classification, conditioning on $Y$ is straightforward because $Y$ belongs to a small set of discrete classes. In regression, because the true target $Y$ is continuous, conditioning on it requires evaluating an infinite number of slices (Perez-Suay et al.; Wang et al.).

To bypass the computational nightmare of conditioning on infinite points, Perez-Suay et al. leveraged kernel methods and cross-covariance operators within Reproducing Kernel Hilbert Spaces (RKHS). Instead of checking independence at every possible value of $Y$, they mapped the continuous variables into a high-dimensional feature space where non-linear dependencies can be measured globally through linear operators. This allowed them to enforce Equalized Odds via a closed-form regularization penalty, avoiding the highly unstable and expensive iterative optimization loops common in adversarial deep learning.

Alternatively, Romano et al. addressed Equalized Odds by introducing "fair dummies." They designed a resampling framework that generates synthetic versions of the sensitive attributes that are, by construction, conditionally independent of the true outcome $Y$. By forcing the model to be unable to distinguish between the real sensitive attributes and these fair dummies during training, they created a highly flexible, differentiable penalty that enforces Equalized Odds for any arbitrary regression architecture.

Reducing Complexity to Known Problems

Rather than designing entirely new optimization paradigms from scratch, a powerful alternative strategy is to reduce fair regression to standard, well-understood machine learning problems.

Agarwal et al. achieved this by discretizing continuous fairness constraints using the cumulative distribution function (CDF). By bounding the differences between group CDFs over a finite grid of points, they transformed the infinite-dimensional fair regression problem into a saddle-point optimization problem. This mathematical formulation allows the problem to be solved as a sequence of standard, cost-sensitive regression tasks, giving practitioners a plug-and-play framework that inherits the theoretical guarantees of traditional empirical risk minimization.

Taking a different reduction path, Narasimhan et al. focused on pairwise fairness. In many scoring and ranking applications, the absolute continuous value is less important than the relative ordering of items. The authors showed that by framing regression as a series of pairwise comparisons (i.e., predicting whether sample $i$ should be scored higher than sample $j$), the continuous problem reduces to a binary classification task. This allowed them to seamlessly port classic classification fairness definitions, like Equal Opportunity, directly into complex continuous scoring systems.

Fairness in Uncertainty

Finally, fairness in continuous spaces extends beyond point predictions ($\hat{Y}$) to the model's expression of uncertainty. If a model predicts a risk score or salary with tight, highly reliable confidence intervals for one demographic group, but provides wide, uninformative intervals for another, the system exhibits a profound fairness defect.

Group A: [-------|-------]  (High Uncertainty / Wide Interval)
Group B:     [---|---]      (Low Uncertainty / Narrow Interval)
             $50k Salary

To address this, Wang et al. introduced Equal Opportunity of Coverage (EOC). Instead of equalizing the predicted value itself, EOC mandates that the prediction intervals (e.g., a 95% confidence window) are equally reliable and uniform across all demographic groups. By developing Binned Fair Quantile Regression (BFQR), the authors provided a distribution-free post-processing method that ensures individuals with similar true outcomes receive equally informative and trustworthy bounds on their predictions, regardless of their protected attributes.

Key Foundational Work

Prediction with Model-Based Neutrality (2013) — contributed 2 concept elements to this direction.
Nonconvex Optimization for Regression with Fairness Constraints (2018) — contributed 1 concept elements to this direction.
Hunting for Discriminatory Proxies in Linear Regression Models (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_fair_regression_3cd676fc,
  title = {fair regression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fair-regression/rdr_3cd676fcdfad4e1d98c6cb3de6dacc23},
  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_fair_regression_3cd676fc_v1,
  title = {fair regression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fair-regression/rdr_3cd676fcdfad4e1d98c6cb3de6dacc23/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake