try to make the acceptance rates for males and females equal, we will introduce a
reverse discrimination, as illustrated in (b). In such case the acceptance rates will be
equal ( 8 =
38% for both); however, females will be privileged, as they will require
a lower experience level than male to qualify for an interview.
We discussed the issues of explainable and illegal discrimination in decision mak-
ing. In the next section we present two techniques to train classifiers with an aim to
remove only the illegal discrimination from future decision making.
Removing the Illegal Discrimination When Training a
In order to ensure that the built classifier is free from illegal discrimination one
needs to control two constraints. Using the university admission example, the two
conditional non-discrimination constraints are as follows:
1. acceptance rates of males and females within each program need to be equal
(although programs may have different acceptance rates, e.g., medicine 20% and
computer science 40%);
2. acceptance rates within each program need to be consistent with the original data,
i.e., even if the acceptance rate to medicine for males and females become equal,
the acceptance rate to medicine should not artificially decreased, e.g., to 10%.
The second condition is necessary to prevent the redlining effect ,asdiscussedinthe
We present two techniques (Zliobaite et al., 2011) for removing illegal discrimina-
tion that modify labels of the historical data so that the historical data is no longer
discriminatory, i.e., it satisfies the two conditional non-discrimination constraints.
The techniques work as follows. First, the 'correct' acceptance rates to each pro-
gram need to be computed. Next, the acceptances of some of the individuals in the
historical data are modified in such a way that the acceptance rates within each
program are 'correct'. The classifiers trained on the modified data, which does not
contain illegal discrimination, are expected to produce decisions that would not con-
tain illegal discrimination. The two presented techniques differ in a way how they
modify the historical data.
The local massaging within every group, defined by a unique value of the explana-
tory attribute 4 , modifies the values of labels until the historical data contains no
E.g., one group will be formed of all students that applied to medicine, the other group
will consists of students that applied to computer science.