DAAR: A Discrimination-Aware Association Rule Classifier for Decision Support

Undesirable correlations between sensitive attributes (such as race, gender or personal status) and the class label (such as recruitment decision and approval of credit card), may lead to biased decision in data analytics. In this paper, we investigate how to build discrimination-aware models even when the available training set is intrinsically discriminating based on the sensitive attributes. We propose a new classification method called Discrimination-Aware Association Rule classifier (DAAR), which integrates a new discrimination-aware measure and an association rule mining algorithm. We evaluate the performance of DAAR on three real datasets from different domains and compare DAAR with two non-discrimination-aware classifiers (a standard association rule classification algorithm and the state-of-the-art association rule algorithm SPARCCC), and also with a recently proposed discrimination-aware decision tree method. Our comprehensive evaluation is based on three measures: predictive accuracy, discrimination score and inclusion score. The results show that DAAR is able to effectively filter out the discriminatory rules and decrease the discrimination severity on all datasets with insignificant impact on the predictive accuracy. We also find that DAAR generates a small set of rules that are easy to understand and applied by users, to help them make discrimination-free decisions.