Yijin Ni

We propose a kernel-based equalized statistic to quantify the accuracy-fairness trade-off among independence-, separation-, and calibration-based constraints, identifying the best suited criterion to preserve predictive accuracy.