Database Reference
In-Depth Information
be combined with direct rule protection. When applying both rule generalization
and direct rule protection,
discriminatory
rules are divided into two groups:
•
Discriminatory
rules
r'
for which there is at least one non-redlining PND rule
r
such that
r'
could be an instance of
r
. For these rules, rule generalization is per-
formed unless direct rule protection requires less data transformation (in which
case direct rule protection is used).
•
Discriminatory
rules
r'
such that there is no such PND rule. For these rules, di-
rect rule protection (DTM 1 or DTM 2) is used.
13.4.2 Indirect Discrimination Prevention Methods
The solution proposed in Hajian
et al.
(2011b) to prevent indirect discrimination is
based on the fact that the dataset of decision rules would be free of indirect dis-
crimination if it contained no
redlining
rules. To achieve this, a suitable data trans-
formation with minimum information loss should be applied in such a way that
redlining
rules are converted to
non-redlining
rules. We call this procedure indi-
rect rule protection (IRP).
Table 13.3
Data transformation methods for indirect rule protection
Indirect Rule Protection
DTM 1
~,, ~→~⇒, , ~→ ~
DTM 2
~, , ~→~⇒~, , ~→
In order to turn a
redlining
rule
r:D, B→C
, where
D
is a non
-
discriminatory
itemset that is highly correlated to the discriminatory itemset
A
, into a
non-
redlining
rule based on the indirect discriminatory measure (
elb
), two data trans-
formation methods could be applied, similar to the ones for direct rule protection.
One method (DTM 1) changes the discriminatory itemset in some records (
e.g.
from non-foreign worker to foreign worker in the records of hired people in NYC
city with Zip≠10451) and the other method (DTM 2) changes the class item in
some records (e.g. from “Hire yes” to “Hire no” in the records of non-foreign
worker of people in NYC city with Zip≠10451). Table 13.3 shows the operation of
these two methods. Table 13.3 shows that in DTM 1 some records in the original
data that support the rule
~A, B, ~D → ~C
will be changed by modifying the val-
ue of the discriminatory itemset from
~A
(Sex=Male) into
A
(Sex=Female) in
these records until the redlining rule
r: D, B →C
becomes non-redlining (
i.e.
elb(r) < α
).
With the aim of scoring better in terms of the utility measures pre-
sented in Section 13.5 and 13.6, among the records supporting the above rule, one
should change those with lowest impact on the other (non-redlining) 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
(
e.g.
grant credit) into
C
(
e.g.
deny credit) in these records to make
r
non-redlining
.
Search WWH ::
Custom Search