At the moment border between fair and unfair is arbitrary. To tackle this issue it would be better for fairness_check() and its plot to look at the ratio of metrics and not the difference. This way 80% rule can be a good boundary of fairness. Each metric should be within:
epsilon <= Metric_unprivileged/Metric_privileged <= 1/epsilon
Which would result in default green area being (0.8, 1.25)
At the same time visualization should not change because of the desirable properties of parity_loss
At the moment border between fair and unfair is arbitrary. To tackle this issue it would be better for
fairness_check()
and its plot to look at the ratio of metrics and not the difference. This way 80% rule can be a good boundary of fairness. Each metric should be within:epsilon <= Metric_unprivileged/Metric_privileged <= 1/epsilon
Which would result in default green area being (0.8, 1.25)At the same time visualization should not change because of the desirable properties of
parity_loss