www-ai.cs.tu-dortmund.de/LEHRE/SEMINARE/SS21/TrustworthyAIMachineLearning/zafar2017a.pdf
in- cluding our own prior work [28], proposing methods for de- tecting [10, 21, 23, 25] and removing [9, 10, 13, 16, 17, 23, 28, 29] unfairness when it is defined in terms of disparate treat- ment, disparate [...] boundary, i.e.:
Cov(z, gθ(y,x)) = E[(z − z̄)(gθ(y,x)− ḡθ(y,x))]
≈ 1
N
∑ (x,y,z)∈D
(z − z̄) gθ(y,x), (9)
where the term E[(z− z̄)]ḡθ(x) cancels out since E[(z− z̄)] = 0 and the function gθ(y,x) is defined [...] is not a trivial task, we ap- proximate it through Monte Carlo covariance on the training set (Eq. 9). While this approximation is expected to work well when a reasonable amount of training data is provided …
