Direct and Indirect Discrimination Prevention Methods - Discrimination and Privacy in the Information Society - page 239

protective (i.e. elift(r') < α ). In order to score better in terms of the utility meas-
ures presented in Section 13.5 and 13.6, the changed records should be those
among the ones supporting the above rule that have the lowest impact on the other
(protective) rules. Similar records are also chosen in DTM 2 with the difference
that, instead of changing discriminatory itemsets, the class item is changed from
~ C (grant credit) into C (deny credit) to make r' protective.

Rule Generalization

Rule generalization is another data transformation method for direct discrimina-
tion prevention. It is based on the fact that if each discriminatory rule r': A, B →C
in the database of decision rules was an instance of at least one non-redlining (le-
gitimate) PND rule r: D, B →C where D is a non-discriminatory itemset
( D nDI s ), the dataset would be free of direct discrimination. To formalize this de-
pendency among rules (i.e. r' is an instance of r ), Pedreschi et al. in (Pedreschi et
al. 2009b) say that a PD classification rule r' is an instance of a PND rule r if rule
r holds with the same or higher confidence, namely conf(r: D,B → C) ≥ conf(r':
A,B→C) , and a case (record) satisfying discriminatory itemset A in context B sa-
tisfies legitimate itemset D as well, namely conf(A, B → D) = 1 .
Based on this concept, a data transformation method ( i.e. rule generalization)
could be applied to transform each discriminatory rule r': A, B →C into an
instance of a legitimate rule. Then, rule generalization can be achieved for discri-
minatory rules r' for which there is at least one non-redlining PND rule r by
changing the class item in some records ( e.g. from “Hire no” to “Hire yes” in the
records of foreign and low-experienced people in NYC city). Table 13.2 shows the
function of this method.

Table 13.2 Data transformation method for rule generalization

Rule Generalization

DTM , , ~→⇒, , ~→ ~

Table 13.2 shows that in DTM some records that support the rule A, B, ~D → C
will change by modifying the value of class item from C (e.g. deny credit) into ~
(e.g. grant credit) until discriminatory rule r': A, B →C becomes an instance of a
non-redlining (legitimate) PND rule r: D, B →C . Similar to DRP methods, in or-
der to score better in terms of the utility measures presented in Section 13.5 and
13.6, the changed records should the ones among those supporting the above rule
that have the lowest impact on the other (protective) rules.

Direct Rule Protection and Rule Generalization

Since rule generalization might not be applicable to all discriminatory rules, rule
generalization cannot be used alone for direct discrimination prevention and must