10 ACTIVE FAIRNESS IN ALGORITHMIC DECISION-MAKING

Society increasingly relies on machine learning models for automated decision-making (ADM). Yet, efficiency gains from automation have come paired with concern for algorithmic discrimination that can systematize inequality. Consequently, research on algorithmic fairness has surged, focusing on either post-processing trained models, constraining learning processes, or pre-processing training data. Recent work has proposed optimal post-processing methods that randomize decisions of a fraction of individuals in order to achieve fairness measures related to calibration and error parity. In contrast, the present work proposes an alternative active framework, where, in deployment, a decision-maker adaptively acquires information according to the needs of different groups or individuals, in order to balance disparities in classification performance. We propose two such methods, where information collection is adapted to group- and individual-level needs, respectively. We show on real-world data sets that these can achieve: (1) calibration and single error parity (e.g., equal false-negatives); and (2) parity in both falsepositive and false-negative rates (i.e., equal odds). Moreover, we show that these fairness goals can be achieved at substantially lower information costs while avoiding other limitations of randomized classifiers such as Pareto suboptimality and intragroup unfairness.