The Discovery of Discrimination - Discrimination and Privacy in the Information Society - page 98

Fig. 5.2 Racial segregation in New York City, based on Census 2000 data (Fischer, 2011).
One dot for each 500 residents. Red dots are Whites, blue dots are Blacks, green dots are
Asian, orange dots are Hispanic, and yellow dots are other races.

even if each person changes his residence only if less than 30% of his neighbors
are of his same race. That's why so many urban territories world-wide, in absence
of social restrictions or incentives, developed a structure such that depicted in Fig-
ure 5.2; in turn, this explains why denying credit or benefits on the basis of residence
- drawing a red line on the border of an urban neighborhood - is often an indirect
way to discriminate on the basis of race. Let us consider an example of inference in
the context of redlining inspired by the Hussein vs Saints Complete House Furniture
case reported by (Makkonen, 2006), albeit the numbers reported here are fictious.
Assume that a Liverpool furniture store refuses to consider 99% of applicants to
a job from a particular postal area ZIP =1234 which had a high rate of unemploy-
ment. The extracted classification rule ZIP =1234, CITY =L IVERPOOL → APP = NO

with confidence 99 is apparently neutral with respect to race discrimination.

Assume also that the average refusal rate in the Liverpool area is much lower, say
9%. With our notation, the rule CITY =L IVERPOOL

γ = 0 .

→ APP = NO has then confi-

dence p 09. Assume now to know, e.g., from census background knowledge,

that 80% of the population in the postal area ZIP =1234 is black, i.e., that the area
is mainly populated by minorities. In formal terms, the association rule ZIP =1234,
CITY =L IVERPOOL

= 0 .

8. It is now legitimate
to ask ourselves whether from such rules, one can conclude a form of redlining,
namely the use of ZIP =1234 as a proxy for excluding blacks from a benefit (accept-
ing the side effect of possibly excluding some whites from the same neighborhood).
Formally, we want to check whether the extended lift of:

→ RACE = BLACK has confidence β = 0 .

( ZIP =1234, RACE = BLACK ), CITY =L IVERPOOL →

APP = NO

( )

is particularly high, where the PD itemset A is ZIP =1234, RACE = BLACK , denoting
blacks living in the area, and the context B is CITY =L IVERPOOL , denoting that the