Algorithmic Fairness for Continuous Sensitive Attributes

This direction is split into other directions: underestimation via attribute binarization, continuous attribute fairness. Its last version stays readable and citable here.

Version 1 of 1186 papers212 conceptsfrozen 2026-10-04

Most algorithmic fairness research operates within a "binary bubble." We often assume that sensitive attributes—like gender or race—are discrete, mutually exclusive categories. However, real-world demographic factors like age, income, credit scores, and geographic coordinates exist on a continuous spectrum.

When we force these continuous variables into discrete bins to fit traditional fairness frameworks, we encounter "border discrimination." An individual just above an arbitrary age threshold might be treated entirely differently than someone just below it, even if their profiles are nearly identical. To solve this, a growing body of research is moving toward "discretization-free" fairness, developing methods that treat sensitive attributes as the fluid, numerical values they actually are.

The Failure of Discretization

The standard approach to fairness involves partitioning a dataset into groups (e.g., "Young" vs. "Old") and ensuring a metric like Demographic Parity is equal across them. As Edward Raff, Jared Sylvester, and Steven Mills argue in Fair Forests: Regularized Tree Induction to Minimize Model Bias, this binarization is fundamentally flawed. It ignores the granular gradients of bias that exist within those bins and creates artificial thresholds where discrimination can hide.

In spatiotemporal contexts, such as the demand prediction models studied by An Yan and Bill Howe in FairST, categorizing geographic regions into discrete groups oversimplifies socioeconomic distributions. This leads to sub-optimal resource allocation because the model cannot "see" the continuous transition of demographics across a city.

Measuring Bias on a Spectrum

If we cannot use simple group-based subtraction to measure bias, how do we quantify it? The intuition is to look for statistical dependence rather than difference in means.

Luca Giuliani and colleagues introduced Generalized Disparate Impact (GeDI) in Generalized Disparate Impact for Configurable Fairness Solutions in ML. GeDI provides a family of indicators specifically for continuous protected attributes. Unlike "black-box" metrics, GeDI is interpretable and robust against the noise often found in small samples, allowing developers to target specific biased dependencies without needing to bin the data.

Another powerful tool is the Hirschfeld-Gebelein-Rényi (HGR) maximal correlation. In A Maximal Correlation Approach to Imposing Fairness in Machine Learning, Joshua Lee and Yuheng Bu demonstrate that HGR can capture complex, non-linear dependencies between model predictions and sensitive variables. By using HGR as a regularizer, they create a smooth, controllable tradeoff between accuracy and fairness that works for both discrete and continuous data.

Enforcing Fairness via Distributional Alignment

Once we can measure bias on a spectrum, the next challenge is enforcing independence. A recurring mechanism for this is distributional alignment—ensuring that the distribution of model outputs is the same regardless of the value of the sensitive attribute.

Vincent Grari and sylvain lamprier pioneered this in Fairness-Aware Neural Rényi Minimization for Continuous Features. They use an adversarial framework to minimize the Rényi correlation, effectively "scrubbing" the influence of continuous variables like age or income from the model's internal representations.

For scenarios where we want to enforce fairness conditional on other features (e.g., ensuring fairness within specific income levels), Mohsen Ghassemi and Alan Mishler propose using Optimal Transport theory. In Auditing and Enforcing Conditional Fairness via Optimal Transport, they introduce algorithms like FairBiT and FairLeap that use the Wasserstein distance to align distributions across complex, continuous conditioning variables. This allows for a more nuanced version of fairness that respects legitimate underlying data structures while removing discriminatory signal.

Integrating Fairness into Model Induction

Beyond post-processing or adversarial training, researchers are baking fairness directly into the learning algorithms themselves.

In the realm of tree-based models, the "Fair Forests" algorithm modifies the splitting criterion of decision trees. Instead of just maximizing information gain for the target label, it penalizes splits that are highly correlated with protected attributes. This ensures that the tree structure itself is built to be "blind" to sensitive numerical gradients.

In Graph Neural Networks (GNNs), where data is interconnected, Huaisheng Zhu and Suhang Wang developed GMMD (Fairness-aware Message Passing for Graph Neural Networks). This framework balances graph smoothness with fairness constraints during the message-passing phase, preventing biased information from propagating through the network even when sensitive features are continuous.

The Path Forward

The shift toward continuous fairness is not just a mathematical refinement; it is a necessity for intersectional fairness. As noted by Jian Kang and Hanghang Tong in InfoFair, individuals exist at the intersection of many identities. When some of those identities are continuous (like age) and others are categorical (like race), traditional group-based auditing fails.

By moving away from discrete bins and toward information-theoretic and distributional alignment tools, the research community is building models that can respect the full, continuous complexity of human identity.

Papers in this version

Showing 60 of 186 papers.

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_algorithmic_fairness_for_continuous_sens_4c9ef811,
  title = {Algorithmic Fairness for Continuous Sensitive Attributes},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/algorithmic-fairness-for-continuous-sensitive-attributes/rdr_4c9ef81104424aafb5df68536e2fe32a},
  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_algorithmic_fairness_for_continuous_sens_4c9ef811_v1,
  title = {Algorithmic Fairness for Continuous Sensitive Attributes},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/algorithmic-fairness-for-continuous-sensitive-attributes/rdr_4c9ef81104424aafb5df68536e2fe32a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}