Explainable and Non-explainable Discrimination in Classification - Discrimination and Privacy in the Information Society - page 159

discrimination. The discrimination model in Section 8.2.3 implies that discrimina-
tion is more likely to affect the objects that are closer to the decision boundary. To
this end, massaging identifies the instances that are close to the decision boundary
and changes the values of their labels to the opposite ones (e.g., positive to neg-
ative or negative to positive). Suppose females have been discriminated as in our
university admission model and the discrimination is reflected in the historical data.
The local massaging identifies a number of females that were almost accepted, and
makes their labels positive, and identifies a number of males that were very likely,
but have not been rejected, and makes their labels negative. To choose the cases
for relabeling, individuals are ordered according to their probability of acceptance
using an internal ranker (a classifier that outputs posterior probabilities), learned on
the training data for each program separately.
The local massaging uses the same principles as the massaging technique
(Kamiran et al., 2010). However, local massaging works on the partitioned data,
within each program separately. In addition, it also modifies and controls the num-
ber of accepted males and females, to ensure no redlining. The procedure for local
massaging is illustrated in Figure 8.6.
Fig. 8.6 Local massaging
Local Preferential Sampling
The preferential sampling technique does not modify the training instances or la-
bels, instead it modifies the composition of the training set. It deletes and duplicates
training instances so that the labels of new training set contain no discrimination.
Following the discrimination model where the discrimination is more likely to
affect the individuals that are closer to the decision boundary, the preferential sam-
pling deletes the 'wrong' cases that are close to the decision boundary and dupli-
cates the cases that are 'correct' and close to the boundary. The cases are selected
using a ranker learned in the same way as in the local massaging. In the university
example the local preferential sampling will delete a number of males that were al-
most rejected and duplicate the males that were almost accepted. It will also delete
a number of females that were almost accepted and duplicate the females that were
almost rejected.