Better Decision Support through Exploratory Discrimination-Aware Data Mining: Foundations and Empirical Evidence

Decision makers in banking, insurance or employment mitigate many oftheir risks by telling ‘‘good’’ individuals and ‘‘bad’’ individuals apart. Laws codifysocietal understandings of which factors are legitimate grounds for differentialtreatment (and when and in which contexts)—or are considered unfair discrimi-nation, including gender, ethnicity or age. Discrimination-aware data mining(DADM) implements the hope that information technology supporting the decisionprocess can also keep it free from unjust grounds. However, constraining datamining to exclude a fixed enumeration of potentially discriminatory features isinsufficient. We argue for complementing it with exploratory DADM, where dis-criminatory patterns are discovered and flagged rather than suppressed. This articlediscusses the relative merits of constraint-oriented and exploratory DADM from aconceptual viewpoint. In addition, we consider the case of loan applications toempirically assess the fitness of both discrimination-aware data mining approachesfor two of their typical usage scenarios: prevention and detection. Using MechanicalTurk, 215 US-based participants were randomly placed in the roles of a bank clerk(discrimination prevention) or a citizen / policy advisor (detection). They weretasked to recommend or predict the approval or denial of a loan, across threeexperimental conditions: discrimination-unaware data mining, exploratory, andconstraint-oriented DADM (eDADM resp. cDADM). The discrimination-aware toolsupport in the eDADM and cDADM treatments led to significantly higher pro-portions of correct decisions, which were also motivated more accurately. There issignificant evidence that the relative advantage of discrimination-aware techniquesdepends on their intended usage. For users focussed on making and motivating their decisions in non-discriminatory ways, cDADM resulted in more accurate and lessdiscriminatory results than eDADM. For users focussed on monitoring for pre-venting discriminatory decisions and motivating these conclusions, eDADM yieldedmore accurate results than cDADM.