Group fairness from two tables
Group A = 0
Group A = 1
Hover over a metric to highlight the cells it uses. Rates like TPR and FPR condition on the true label Y (a column of the table); PPV conditions on the prediction Ŷ (a row).

Randomized post-processing: derived classifier Ỹ with p(y, a) = P(Ỹ = 1 | Ŷ = y, A = a)

p(1, 0): keep Ŷ = 1 in group 0
p(0, 0): flip Ŷ = 0 to 1 in group 0
p(1, 1): keep Ŷ = 1 in group 1
p(0, 1): flip Ŷ = 0 to 1 in group 1