www-ai.cs.tu-dortmund.de/de/LEHRE/SEMINARE/WS2021/TrustworthyAIMachineLearning/literature/zafar2019.pdf
and assign a 2-dimensional user feature vector per label by drawing samples from
p(x|y = 1, β) = βN ([2; 2], [5 1; 1 5]) + (1− β)N ([−2;−2], [10 1; 1 3])
p(x|y = −1, β) = βN ([4;−4], [4 4; 2 5]) + (1− β)N [...] on):
p(x|z = 0, y = 1) = N ([2, 2], [3, 1; 1, 3])
p(x|z = 1, y = 1) = N ([2, 2], [3, 1; 1, 3])
p(x|z = 0, y = −1) = N ([1, 1], [3, 3; 1, 3])
p(x|z = 1, y = −1) = N ([−2,−2], [3, 1; 1, 3]).
17
Zafar, Valera [...] deem classifiers C2 and C3 to be unfair due to disparate treatment since C2’s (C3’s) decisions for Male 1 and Female 1 (Male 2 and Female 2) are different even though they have the
2. Technically, the …