Active Fairness in Algorithmic Decision-Making
Society increasingly relies on machine learning models for automated decision making across domains like criminal justice, human resources, medical diagnosis, social work, credit scoring and insurance. Efficiency gains from such automation have come paired with increasing concern for how unfair models can deepen inequality. Intuitively, a predictive model is considered group-fair if it yields equal error rates across population sub-groups-e.g., gender or racial groups. Recently, prominent work in algorithmic fairness has focused on either a) optimal post-processing adjustments to black-box models (Hardt et. al 2016); or b) given a labeled dataset, enforcing fairness through clever optimization constraints within supervised learning algorithms (Zemel et. al 2016). In contrast, we propose an active fairness approach, where, for each test subject, the system adaptively collects information (features) until a confidence threshold is met, upon which a decision is made. We propose two greedy algorithms for efficient active fairness, based on random forests and Gaussian mixtures, respectively. We show that active fairness: i) can balance both false positive and false negative rates (in contrast to post-processing methods); ii) achieve fairness for all sub-groups, without requiring to specify them, nor requiring sensitive labels during training nor testing; and iii) reduce information cost in production.
Authors: Alejandro Noriega, Michiel Bakker, Bruke Kifle, and Alex Pentland