PAC-Bayesian Generalization Guarantees for Fairness on Stochastic and Deterministic Classifiers

Researchers have developed a PAC-Bayesian framework that provides mathematical generalization guarantees for fairness in both stochastic and deterministic classifiers. The approach covers a broad class of fairness measures expressed as risk discrepancies and leads to self-bounding algorithms that directly optimize the trade-off between prediction accuracy and fairness constraints. Empirical evaluation with three classical fairness measures demonstrates both the framework's usefulness and the tightness of the derived bounds.