Database Reference
In-Depth Information
The difference between the DRP and IRP methods shown in Tables 1 and 3 is
about the set of records chosen for transformation. As shown in Table 3, in IRP
the chosen records should not satisfy the
D
itemset (chosen records are those with
~A, B,~D→~C
), whereas DRP does not care about
D
at all (chosen records are
those with
~A, B →~
C
).
13.5 Measuring Discrimination Removal
Discrimination prevention methods should be evaluated based on two aspects: dis-
crimination removal and data quality. We deal with the first aspect in this section:
how successful the method is at removing all evidence of direct and/or indirect
discrimination from the original dataset. To measure discrimination removal, four
metrics were proposed in Hajian
et al.
(2011a and 2011b) and Hajian and Domin-
go-Ferrer (2012):
•
Direct Discrimination Prevention Degree (DDPD).
This measure quantifies
the percentage of
discriminatory
rules that are no longer
discriminatory
in the
transformed dataset.
Direct Discrimination Protection Preservation
(DDPP)
. This measure quan-
tifies the percentage of the
protective
rules in the original dataset that remain
protective
in the transformed dataset.
•
Indirect Discrimination Prevention Degree (IDPD).
This measure quantifies
the percentage of
redlining rules
that are no longer
redlining
in the transformed
dataset.
•
Indirect Discrimination Protection Preservation (IDPP).
This measure
quantifies the percentage of
non-redlining
rules in the original dataset that re-
main
non-redlining
in the transformed dataset.
•
Since the above measures are used to evaluate the success of the proposed
methods in direct and indirect discrimination prevention, ideally their value should
be 100%.
13.6 Measuring Data Quality
The second aspect to evaluate discrimination prevention methods is how much in-
formation loss (
i.e.
data quality loss) they cause. To measure data quality, two me-
trics are proposed in Verykios and Gkoulalas-Divanis (2008):
•
Misses Cost
(MC). This measure quantifies the percentage of rules among
those extractable from the original dataset that cannot be extracted from the
transformed dataset (side-effect of the transformation process).
•
Ghost Cost
(GC). This measure quantifies the percentage of the rules among
those extractable from the transformed dataset that were not extractable from
the original dataset (side-effect of the transformation process).
Search WWH ::
Custom Search