www-ai.cs.tu-dortmund.de/LEHRE/SEMINARE/WS2021/TrustworthyAIMachineLearning/literature/zafar2019.pdf
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 ([−4; 6], [6 2; 2 3])
where β ∈ {0, 1} is sampled from Bernoulli(0.5). Then, we generate [...] comparison with Eq. (4.10), Eq. (4.11) involves a tighter set of constraints—it aims to bound individual users’ losses as opposed to the aggregate loss over all the users. As a result, Eq. (4.11) could lead [...] values: φ = π/4 and
16
Fairness Constraints: A Flexible Approach for Fair Classification
Acc=0.94; AR=0.28:0.66(π/4),0.11:0.85(π/8)
(a) Unconstrained
Acc=0.83; AR=0.55:0.57
(b) φ = π/4
Acc=0.56; AR=0 …
