Generalized Disparate Impact for Configurable Fairness Solutions in ML

The paper addresses a gap in fairness measurement for continuous protected attributes (like age or income, rather than categorical ones like gender). The existing Hirschfeld-Gebelein-Renyi (HGR) indicator was the only available option, but the authors demonstrate it has significant limitations in how it interprets and explains fairness issues.

New fairness metrics offer better interpretability and transparency for stakeholders. Unlike the HGR indicator, the authors' proposed family of indicators are "fully interpretable and transparent," which means organizations can more clearly explain to affected individuals and regulators why certain decisions were made and how fairness was evaluated.

The new approach is more robust with real-world data. The proposed indicators are "robust over finite samples," addressing a practical problem: fairness metrics that work perfectly in theory often fail when applied to actual datasets of limited size, which is common in real business applications.

Organizations can now customize fairness constraints to their specific context. The approach allows "fine-grained constraints to permit certain types of dependence and forbid others selectively," meaning different industries or applications can define fairness in ways appropriate to their specific risks and values rather than applying one-size-fits-all rules.

This expands fairness options beyond what was previously available. By providing multiple complementary indicators for continuous protected attributes, the work gives practitioners more tools to identify and address different types of unfairness in machine learning systems.