Fig. 6. Comparative performance of different methods on sex and handedness.

Four methods have been tested: Zhang et al., Ganin et al., MDANN with 0 adversarial components (AC), and MDANN with 2 ACs, across two sensitive features: (a) Sex and (b) Handedness. The three performance metrics are represented by different line styles and markers: Accuracy (ACC) is shown in blue with circular markers and a solid line, Area Under the Receiver Operating Characteristic Curve (AUC) is depicted in green with square markers and a dashed line, and Disparate Impact (DI) is illustrated in red with triangular markers and a dash-dot line. A horizontal black dashed line at y=1 serves as a reference for the DI metric, where values closer to 1 indicate enhanced fairness. The results highlight the superior performance of MDANN, particularly the variation with 2 ACs, in terms of both predictive accuracy and fairness.