Fair AI Representations for Continuous Sensitive Attributes.

Achieving fairness in AI models, especially when dealing with sensitive attributes like age or income that are continuous, typically requires ensuring that the learned data representation is statistically independent of these attributes. Current methods often do this by calculating and averaging discrepancies between conditional and marginal probability distributions for each sensitive value, which can be computationally intensive and statistically inefficient due to the need for nonparametric surrogates.This research introduces a more efficient approach: evaluating independence using a single "joint discrepancy" between the joint probability distribution of the representation and the sensitive attribute, and the product of their marginals. This method avoids the need for conditional law estimation. Using the Hilbert-Schmidt independence criterion (HSIC) as an example, the authors demonstrate that this joint-distribution route offers a closed-form estimator with a faster convergence rate ($O(n^{-1/2})$) compared to the $O(n^{-2/5})$ rate of conditional-route estimators. The resulting algorithm, FRHSIC, achieves comparable fairness-accuracy trade-offs while significantly reducing training time per epoch.