The Discovery of Discrimination - Discrimination and Privacy in the Information Society - page 101

important to assess and to monitor the application of affirmative actions. In our ap-
proach, affirmative actions can be unveiled by proceedings in a similar way as for
discriminatory actions. The basic idea is to search, either directly or indirectly, for
a -discriminatory PD rules of the form:
A BENEFIT = GRANTED

i.e., where the consequent consists of granting a benefit (a loan, a school admission,
a job, etc.). Rules of this form with a value of the discrimination measure greater
than a fixed threshold highlight contexts B where the disadvantaged group A was
actually favored.
Once again, consider our running example dataset. By ranking classification rules
of the form A

, B →

CLASS = GOOD accordingly to their extended lift measure, we
found near the top positions the following:
AGE = GT 52, JOB = UNEMPLOYED

, B →

CLASS = GOOD
with an extended lift of 1.39. The rule can be interpreted as follows: among those un-
employed, people older than 52 had 1.39 times the average chance of being granted
the requested credit. This could be the case, for instance, of some affirmative actions
supporting economic initiatives of unemployed older people.

→

5.7 The DCUBE Tool

The various concepts and analyses so far discussed, originally implemented as
stand-alone programs for achieving the best performances, have been re-designed
around an Oracle database, used to store extracted rules, and a collection of func-
tions, procedures and snippets of SQL queries that implement the various legal rea-
sonings for discrimination analysis. The resulting implementation, called DCUBE
(Discrimination Discovery in Databases) (Ruggieri et al., 2010), can be accessed
and exploited by a wider audience if compared to a stand-alone monolithic applica-
tion. In fact, SQL is the dominant query language for relational data, with database
administrators already mastering issues such as data storage, query optimization,
and import/export towards other formats. Discrimination discovery is an interac-
tive and iterative process, where analyses assume the form of deductive reasoning
over extracted rules. An appropriately designed database, with optimized indexes,
functions and SQL query snippets, can be welcome by a large audience of users, in-
cluding owners of socially-sensitive decision data, government anti-discrimination
analysts, technical consultants in legal cases, researchers in social sciences, eco-
nomics and law. Typical discrimination discovery questions that DCUBE is able to
answer include:
Direct discrimination discovery: “How much have women been under-represen-
ted in obtaining the loan?” or “List under which conditions blacks were suffering
an extended lift higher than 1.8 in our recruitment data” . DCUBE comes with
all of the legally-grounded measures from Figure 5.1 predefined. The user can
adopt any of them or, even, she can easily define new measures over a 4-fold
contingency table by adding methods to an Oracle user defined data type.