Most foundational work in algorithmic fairness assumes that protected characteristics belong to neat, discrete categories—such as binary gender or distinct racial groups. However, many real-world sensitive attributes exist on a continuous spectrum: age, income, credit scores, or geographic coordinates.
As noted by Luca Oneto et al. in General Fair Empirical Risk Minimization, forcing these numerical spectrums into discrete "bins" or "buckets" introduces severe limitations. Discretization creates arbitrary threshold effects: if an algorithm treats age 29 and age 30 as belonging to different buckets, it introduces artificial discontinuities. Furthermore, binning destroys the natural ordering and information density of the data, potentially masking non-linear biases that fluctuate smoothly across the spectrum. To resolve this, a growing body of research focuses on enforcing fairness directly over continuous domains.
Redefining Demographic Parity for Spectrums
Traditional demographic parity requires a model's predictions $\hat{Y}$ to be statistically independent of a discrete protected attribute $A$, meaning $P(\hat{Y} \mid A = a) = P(\hat{Y})$. When $A$ is continuous, calculating these conditional probabilities becomes problematic due to data sparsity—any exact value of income or age might only appear once in a dataset.
To bridge this gap, Zhimeng Jiang et al. proposed Generalized Demographic Parity for Group Fairness, which measures the divergence between local conditional prediction averages and the global average across the entire spectrum. Similarly, Luca Giuliani et al. introduced Generalized Disparate Impact (GeDI) in Generalized Disparate Impact for Configurable Fairness Solutions in ML, providing an interpretable framework to track and eliminate biased dependencies without relying on categorical group boundaries. By evaluating fairness as a continuous function, these metrics ensure that adjustments vary smoothly based on the exact value of the sensitive attribute.
Information-Theoretic and Non-Linear Correlation Mitigation
How do we enforce independence between a model's output and a continuous attribute during training? Simple linear correlation (like Pearson's $r$) is insufficient because bias can be highly non-linear. A model might be uncorrelated with age overall, yet highly biased against both very young and very old individuals.
A powerful solution lies in the Hirschfeld-Gebelein-Rényi (HGR) maximal correlation coefficient. The HGR coefficient looks for the worst-case non-linear transformations of both the model's predictions and the sensitive attribute that maximize their linear correlation. By minimizing this coefficient, models can decouple complex, non-linear dependencies.
Jeremie Mary et al. demonstrated in Fairness-Aware Learning for Continuous Attributes and Treatments that the HGR coefficient can be upper-bounded by a differentiable $\chi^2$-divergence penalty, enabling fairness-aware deep learning via standard gradient descent. Vincent Grari expanded on this in Fairness-Aware Neural Rényi Minimization for Continuous Features, utilizing an adversarial neural network framework to directly minimize HGR and eliminate non-linear leakage for continuous variables. This adversarial approach to learning unbiased representations is echoed by Marcin Detyniecki et al. in Learning Unbiased Representations via Rényi Minimization, forcing neural networks to learn internal features that share zero mutual information with the protected spectrum.
Kernel Methods and Statistical Independence
An alternative to adversarial optimization—which can be notoriously unstable—is mapping the continuous attributes into a Reproducing Kernel Hilbert Space (RKHS). In this infinite-dimensional space, complex statistical dependencies can be measured globally using kernel operators.
Tennison Liu et al. developed FairCOCCO in Practical Approaches for Fair Learning with Multitype and Multivariate Sensitive Attributes, which leverages kernel-based cross-covariance operators to constrain multiple continuous or mixed sensitive attributes simultaneously. By using kernel functions, the framework preserves the natural ordering and information density of attributes like age or credit limits while avoiding artificial threshold effects.
To bypass adversarial training in fair representation learning, Yongdai Kim et al. introduced FREM in Fair Representation Learning for Continuous Sensitive Attributes using Expectation of Integral Probability Metrics. FREM applies kernel smoothing to continuous sensitive attributes, computing an Expectation of Integral Probability Metrics (EIPM) via Maximum Mean Discrepancy (MMD). This approach mathematically guarantees that representation spaces remain fair across a continuous spectrum without losing structural data continuity.
Geometric Projections and Pre-Processing
Fairness can also be achieved by modifying the data geometry before it reaches a model, or by projecting representations into a "bias-free" subspace.
Shixiang Zhu et al. proposed a model-agnostic pre-processing algorithm called Orthogonal to Bias (OB) in Counterfactual Fairness Through Transforming Data Orthogonal to Bias. OB ensures counterfactual fairness by transforming non-sensitive features so that they are geometrically orthogonal to continuous sensitive variables. For post-hoc representation scrubbing, Felix Störck et al. extended null-space projections to non-linear spaces in Extending Fair Null-Space Projections for Continuous Attributes to Kernel Methods, allowing developers to mathematically "erase" continuous sensitive information from kernel matrices.
Expanding the Scope: Spatial Fairness
The transition from discrete categories to continuous domains has unlocked solutions for complex, multi-dimensional problems like spatial fairness. Traditional geographic fairness relies on fixed boundaries like zip codes, which are prone to statistical fragility and gerrymandering.
As argued by Yiqun Xie et al. in Fairness by "Where": A Statistically-Robust and Model-Agnostic Bi-level Learning Framework, treating geographic space as a continuous distribution of potential regional partitions prevents the arbitrary boundary effects of static categorization. By modeling location as a continuous spectrum rather than a discrete category, these frameworks ensure equitable model outcomes across continuous geographic landscapes.
Key Foundational Work
Prediction with Model-Based Neutrality (2015) — contributed 1 concept elements to this direction.
Two-stage Algorithm for Fairness-aware Machine Learning (2017) — contributed 1 concept elements to this direction.
Fair Forests: Regularized Tree Induction to Minimize Model Bias (2017) — contributed 1 concept elements to this direction.