Database Reference
In-Depth Information
rules, so that no unfair decision rule can be mined from the transformed data. As
part of this effort, the metrics that specify which records should be changed, how
many records should be changed and how those records should be changed during
data transformation are developed.
There are some assumptions common to all methods in this section. First, we as-
sume the class attribute in the original dataset DB to be binary ( e.g. denying or
granting credit). Second, we obtain the database of discriminatory and redlining
rules as output of a discrimination measurement (discovery) phase based on meas-
ures proposed in Pedreschi et al. (2008) and Pedreschi et al. (2009a); discrimination
measurement is performed to identify discriminatory and redlining rules (based on
the work in Chapter 5); then a data transformation phase is needed to transform the
data in order to remove all evidence of direct or indirect discriminatory biases asso-
ciated to discriminatory or redlining rules. Third, we assume the discriminatory
itemsets ( i.e. A ) and the non-discriminatory itemsets ( i.e. D ) to be categorical.
13.4.1 Direct Discrimination Prevention Methods
The proposed solution to prevent direct discrimination is based on the fact that the
dataset of decision rules would be free of direct discrimination if it only contained
PD rules that are protective or PD rules that are instances of at least one non-
redlining (legitimate) PND rule. Therefore, a suitable data transformation with
minimum information loss should be applied in such a way that each discrimina-
tory rule either becomes protective or an instance of a non-redlining PND rule.
We call the first direct rule protection and the second one rule generalization.
Direct Rule Protection (DRP)
In order to convert each discriminatory rule r': A, B →C , where A is a discriminato-
ry itemset ( A
nDI s ) ), into a protec-
tive rule, two data transformation methods (DTM) could be applied. One method
(DTM 1) changes the discriminatory itemset in some records ( e.g. gender changed
from male to female in the records with granted credits) and the other method (DTM
2) changes the class item in some records ( e.g. from grant credit to deny credit in the
records with male gender). Table 13.1 shows the operation of these two methods.
DI s ) and B is non-discriminatory itemset (B
Table 13.1 Data transformation methods for direct rule protection
Direct Rule Protection
DTM 1 ~ , →~ , → ~
DTM 2 ~ , →~ ~ ,
Table 13.1 shows that in DTM 1 some records supporting rule ~ , ~ will
be changed by modifying the value of the discriminatory itemset from ~ A
(Sex=Male) to A (Sex=Female) until discriminatory rule r': A, B →C becomes
 
Search WWH ::




Custom Search