Algorithmic fairness has traditionally operated in a world of discrete boxes. Standard fairness metrics—such as demographic parity or equalized odds—typically require splitting individuals into distinct, categorical groups (e.g., "privileged" versus "non-privileged").
However, real-world demographic and socioeconomic data rarely fits neatly into binary or categorical buckets. As Luca Oneto et al. point out in General Fair Empirical Risk Minimization, many critical sensitive attributes are inherently continuous, such as age, income, credit limits, or geographic population densities.
To apply classical fairness interventions to these continuous attributes, practitioners frequently resort to discretization—the practice of carving a continuous spectrum into arbitrary bins (e.g., splitting age into "under 30" and "30 and over"). While this binarization makes traditional metrics easy to compute, it introduces severe statistical artifacts that undermine the very fairness goals it attempts to achieve.
The Hidden Costs of Discretization
When we force a continuous spectrum into discrete categories, we introduce three core pathologies: information loss, boundary effects, and the systematic underestimation of bias.
1. Information Loss and Intra-Bin Variance
Discretization treats everyone within a single bin as completely identical and everyone across a boundary as fundamentally different. Zhimeng Jiang et al. demonstrate in Generalized Demographic Parity for Group Fairness that this process completely destroys the natural ordering of the data. By flattening the variance within a bin, the model remains blind to unfair disparities occurring inside that bucket.
2. Boundary Effects and Subgroup Fragmentation
Arbitrary thresholds create severe discontinuities. In Fairness-Aware Learning for Continuous Attributes and Treatments, Jeremie Mary et al. observe that a model can appear perfectly fair according to a binarized metric while maintaining significant, localized bias for individuals situated right next to each other on opposite sides of a bin boundary.
Furthermore, these boundaries can fragment cohesive demographic patterns. José María Luna et al. show in Exceptional in so Many Ways - Discovering Descriptors That Display Exceptional Behavior on Contrasting Scenarios that artificial binning splits statistically significant, vulnerable subgroups across multiple intervals, making them invisible to auditing algorithms. This fragmentation directly impairs our ability to diagnose dataset coverage gaps, a vulnerability highlighted by Zhongjun Jin et al. in Assessing and Remedying Coverage for a Given Dataset.
3. Underestimation of Discrimination
Perhaps most dangerously, binarization actively masks the true extent of algorithmic discrimination. Yijun Bian et al. in Do Existing Fairness Measures Suffice? Assessing Discrimination in Algorithmic Decision-Making prove that simplifying multi-valued or continuous attributes into binary categories systematically underestimates discrimination, hiding deep disparities under a veneer of mathematical compliance.
| Preprocessing Approach | Mathematical Impact | Fairness Consequence |
|---|---|---|
| Binarization / Binning | Destroys local continuity; introduces step-function thresholds. | Appears fair on paper; masks localized bias near boundaries and within bins. |
| Continuous Modeling | Preserves data topology, ordering, and exact statistical dependencies. | Guarantees local fairness across the entire spectrum. |
Why Traditional Metrics Fail on the Spectrum
To understand why we cannot simply "fix" binning by adding more bins, consider how fairness constraints interact with optimization.
If you create too few bins, you suffer from the massive information loss described above. If you create too many bins to capture the nuances, you over-partition the data. As Luca Oneto et al. demonstrate, over-partitioning severely restricts the optimization space during empirical risk minimization, causing a drastic, unnecessary degradation in predictive accuracy.
Even when discretization happens at the very end of a pipeline—such as converting continuous eigenvectors into discrete cluster assignments in graph networks—it introduces mathematical complications. Dongqi Fu et al. note in Fairness-Aware Clique-Preserving Spectral Clustering of Temporal Graphs that establishing a theoretical bound for the fairness error introduced by final discretization steps remains an open mathematical challenge.
Similarly, when trying to adapt multiclass or multi-valued datasets to binary frameworks, Ibrahim Alabdulmohsin et al. (A Reduction to Binary Approach for Debiasing Multiclass Datasets) and Hyungrok Do et al. (Fair Generalized Linear Models with a Convex Penalty) emphasize that heuristic binning creates a layer of approximation that fails to capture the true underlying data distribution. Carolin Strobl et al. (An R toolbox for score-based measurement invariance tests in IRT models) show that this approximation frequently misses the exact "break points" where measurement bias actually manifests.
Moving Beyond Bins: Continuous Fairness Frameworks
To resolve these limitations, recent research shifts away from comparing discrete group averages. Instead, it focuses on optimizing smooth statistical independence across the entire continuous distribution.
Traditional Binning:
Continuous Attribute ──> [ Bin 1 ] vs [ Bin 2 ] ──> Compare Discrete Averages (Masks Boundary Bias)
Continuous Fairness:
Continuous Attribute ──> Smooth Statistical Dependence Measure ──> Differentiable Penalty (Preserves Topology)
Several distinct mathematical frameworks have emerged to achieve this without relying on arbitrary thresholds:
Non-Linear Correlation Penalties
Instead of checking if two groups have the same average outcome, we can penalize the overall non-linear correlation between the model's predictions and the continuous sensitive attribute. Jeremie Mary et al. achieve this by leveraging the Rényi maximum correlation coefficient (HGR). By upper-bounding the HGR with a differentiable $\chi^2$-divergence penalty, they enable deep neural networks to smoothly scrub bias during standard mini-batch gradient descent.
Kernel-Based Independence Operators
Another approach maps the continuous attributes into an infinite-dimensional feature space where complex, non-linear dependencies become linear and easily measurable.
- Tennison Liu et al. introduce FairCOCCO in Practical Approaches for Fair Learning with Multitype and Multivariate Sensitive Attributes, utilizing kernel-based cross-covariance operators to penalize dependencies smoothly across multivariate, continuous attributes simultaneously.
- Felix Störck et al. take this a step further in Extending Fair Null-Space Projections for Continuous Attributes to Kernel Methods by directly transforming the kernel matrix to project continuous sensitive information into the null space, effectively "scrubbing" it from regression tasks while preserving predictive utility.
Integral Probability Metrics (IPMs)
To guarantee fairness locally across an entire spectrum, representation learning must account for local continuity. In Fair Representation Learning for Continuous Sensitive Attributes using Expectation of Integral Probability Metrics, Yongdai Kim et al. propose the Expectation of Integral Probability Metrics (EIPM). Combined with a kernel-smoothed estimator, this framework enforces fair representations across continuous attributes without the training instability of adversarial networks or the granular information loss of binning.
Distributionally Robust Optimization (DRO)
When explicit demographic labels are unreliable or continuous distributions are highly complex, fairness can be framed as a minimax problem over latent sub-distributions.
- Siqi Li et al. develop ROME (Robust Mixture Models for Algorithmic Fairness Under Latent Heterogeneity), which automatically discovers latent subgroups hidden within continuous data interactions and applies DRO to optimize for the most disadvantaged, unobserved populations.
- Similarly, Shu Hu et al. demonstrate in Distributionally Robust Survival Analysis: A Novel Fairness Loss Without Demographics that minimizing worst-case error across all sufficiently large subpopulations inherently improves fairness for continuous variables like age, bypassing the need for arbitrary demographic bucketing entirely during training.
By treating fairness as a continuous landscape rather than a collection of discrete territories, these frameworks ensure that protection is applied uniformly, eliminating the blind spots and boundary artifacts of traditional algorithmic auditing.
Key Foundational Work
Assessing and Remedying Coverage for a Given Dataset (2019) — contributed 2 concept elements to this direction.
General Fair Empirical Risk Minimization (2020) — contributed 2 concept elements to this direction.
Exceptional in so Many Ways - Discovering Descriptors That Display Exceptional Behavior on Contrasting Scenarios (2020) — contributed 2 concept elements to this direction.