IM_LR: An approach for Direct and Indirect Discrimination Prevention

Abstract

Discrimination and privacy preservation are major challenges of data mining. Technique based on impact minimization to prevent discrimination has been reported in the literature. The technique computes fitness of generated frequent rules based on their antecedent, a pre-defined threshold and discrimination measure ‘elift’ to modify discriminating rules. This paper deals with a method called ‘IMLR’. IMLR computes fitness of generated frequent rules based on their antecedent (attributes on left hand side of the rule) as well as consequences (class label on right hand side of the rule), a pre-defined threshold and offers selection of desired discrimination measures such as ‘elift’, ‘slift’, ‘olift’ etc. to modify discriminating rules. Experimentation results carried out using two well-known datasets ‘Adult’ and ‘German’ show that IMLR when used with certain discrimination measure provides better results in terms of various performance parameters such as DDPD, DDPP, IDPD, IDPP, Missed cost and Ghost cost when compared with reported technique