Database Reference
In-Depth Information
CLASS = BAD
with an extended lift of 1.58. The rule can be interpreted as follows: among those
applying for loans to buy a new car, female foreign workers had 1.58 times the av-
erage chance of being refused the requested credit. The rule above has a confidence
of 0.277, meaning that female foreign workers asking a loan to buy a new car had
credit denied in 27.7% of cases (precisely, 13 transactions out of 47). The rule for
the generality of applicants:
PURPOSE = NEW CAR
PURPOSE = NEW CAR
CLASS = BAD
has a confidence of 0.175, meaning that people asking a loan to buy a new car had
credit denied in 17.5% of cases.
5.4
Indirect Discrimination Discovery
The EU Directives (see European Union Legislation, 2011; Tobler, 2008) provide
a broad definition of indirect discrimination (also known as systematic discrimi-
nation or disparate impact) as occurring “where an apparently neutral provision,
criterion or practice would put persons of a racial or ethnic origin at a particular
disadvantage compared with other persons”. In other words, the actual result of
the apparently neutral provision is the same as an explicitly discriminatory one.
In our framework, the “actual result” is modeled by a PD rule A
C that is
a -directly discriminatory, while an “apparently neutral provision” is modeled by a
potentially non-discriminatory (PND) rule B
,
B
C , where PD itemsets do not oc-
cur at all. The issue with unveiling indirect discrimination is that the actual result
A
C may be unavailable 4 , e.g., because the dataset does not contain attributes
to denote the potentially discriminated groups. For instance, the information on a
person's race is typically not available and, in many countries, not even collectable.
In our approach to indirect discrimination, the problem consists then of inferring
some PD rule (with a high discrimination measure value) starting from the set of
PND rules, and, possibly, from additional background knowledge. The adjective
potentially non-discriminatory was chosen exactly to underline that, since the rule
does not refer to protected groups, it does not unveil any discriminatory practice in
a direct way. Nevertheless, it could do that indirectly.
A remarkable example is redlining , a form of indirect discrimination that is ex-
plicitly banned in the US (U.S. Federal Legislation, 2011, (b)). As sharply pointed
out in Figure 5.2, racial segregation very often emerges in most cities character-
ized by ethnic diversity: the spatial clustering of a city into racially homogeneous
areas is observed in reality much more often than the dispersion of races into an in-
tegrated structure. We know from Schelling's segregation model (Schelling, 1971)
that a natural tendency to spatial segregation emerges, as a collective phenomenon,
even if each individual person is relatively tolerant and open-minded: in his famous
abstract simulation model, Schelling showed how segregation eventually appears
,
B
4
Otherwise, the technique of Section 5.3 can be adopted to unveil the effects of both direct
and indirect discrimination.
Search WWH ::




Custom Search