Abstract
It is important to guarantee that machine learning algorithms deployed in the real world do not result in unfairness or unintended social consequences. Fair ML has largely focused on the protection of single attributes in the simpler setting where both attributes and target outcomes are binary. However, the practical application in many a real-world problem entails the simultaneous protection of multiple sensitive attributes, which are often not simply binary, but continuous or categorical. To address this more challenging task, we introduce FairCOCCO, a fairness measure built on cross-covariance operators on reproducing kernel Hilbert Spaces. This leads to two practical tools: first, the FairCOCCO Score, a normalised metric that can quantify fairness in settings with single or multiple sensitive attributes of arbitrary type; and second, a subsequent regularisation term that can be incorporated into arbitrary learning objectives to obtain fair predictors. These contributions address crucial gaps in the algorithmic fairness literature, and we empirically demonstrate consistent improvements against state-of-the-art techniques in balancing predictive power and fairness on real-world datasets.
Introduction: Beyond Binary Fairness
Algorithmic fairness research has traditionally operated within a simplified paradigm: protecting a single binary attribute (e.g., gender) against a binary outcome. However, real-world discrimination is rarely so isolated. Legal frameworks, such as those in the U.S., protect numerous classes simultaneously. Protecting these attributes individually is insufficient due to "fairness gerrymandering"—where a model appears fair for individual groups but remains biased against their intersections (e.g., Black women). Furthermore, many sensitive attributes like age or credit limits are continuous; standard metrics often require arbitrary "binning," which discards information and introduces threshold artifacts.
The FairCOCCO Framework
To address these limitations, we propose FairCOCCO (Fairness through Cross-COvariance COnditional operators). By leveraging Reproducing Kernel Hilbert Spaces (RKHS), FairCOCCO maps data into a high-dimensional space to capture complex, higher-order dependencies that linear correlations miss.
The framework provides two primary tools:
- The FairCOCCO Score: A normalized metric (0 to 1) quantifying the dependency of predictions on sensitive attributes.
- The FairCOCCO Regularizer: A differentiable penalty term for model training that balances predictive accuracy with fairness constraints.
Crucially, FairCOCCO is framework-agnostic, supporting various fairness definitions like Demographic Parity (DP) and Equalized Odds (EO) across diverse data types, including tabular, image, and time-series data.
Experimental Evaluation
We evaluated FairCOCCO across 11 diverse datasets, ranging from standard benchmarks like COMPAS and Adult to complex multivariate tasks like Crimes and Communities (C&C) and clinical time-series data from MIMIC-III. This broad evaluation demonstrates the method's versatility in handling both categorical and continuous sensitive attributes.
Key Results and Discussion
Empirical results show that FairCOCCO consistently achieves a superior Pareto frontier compared to state-of-the-art methods like FACL and FARMI. As illustrated in Figure 1, FairCOCCO provides a better trade-off between predictive error (MSE) and fairness, achieving lower bias scores for the same level of accuracy.
On standard benchmarks, FairCOCCO often outperforms existing techniques. Notably, on the COMPAS dataset, it achieved a Difference in Equal Opportunity (DEO) of 0.00 (see Figure 2), effectively eliminating measurable bias while maintaining high accuracy.
In complex scenarios, such as sepsis treatment policy in the MIMIC-III dataset, FairCOCCO successfully reduced bias in clinical interventions without compromising the model's utility in suggesting life-saving treatments.
Conclusion and Future Work
FairCOCCO addresses critical gaps in multivariate and continuous fairness. While the current implementation relies on kernel operations with $O(N^3)$ complexity—mitigated by low-rank approximations—future work will focus on further scaling these operations. Additionally, extending the regularizer to non-differentiable models like Random Forests remains an open area for research.
Summary updated