An Empirical Evaluation of various Discrimination Measures for Discrimination Prevention

Abstract

Discrimination prevention in Data mining has been studied by researchers. Several methods have been devised to take care of both direct and indirect discrimination prevention. In order to prevent discrimination, each of these methods tries to minimize the impact of discriminating attributes by modifying certain discriminating rules. The discriminating rules are identified using certain threshold and discrimination measure such as elift for direct discrimination and elb for indirect discrimination. Performance of these methods are measured and compared in terms discrimination removal using DDPD, DDPP for direct discrimination and IDPD, IDPP for indirect discrimination as well as resultant data quality using MC and GC for both kinds of discrimination.This paper deals with study of use of discrimination measures other than elift such as slift, clift and olift. The empirical evaluation presented here shows that slift provides best overall performance.