Fig. 4. Prediction results addressing the label-imbalance problem.

Three sensitive attributes from ADNI datasets were included: age (with aged over 78 years as the minority group), educated years (with educated over 16 years as the minority group), and handedness (with left-handed as the minority group). Rows represent the prediction metrics with different loss functions: cross-entropy loss and minimax loss. The first three columns show the accuracy of the predictor when predicting sensitive features respectively, and the last column shows the AUC score of the predictor.