On Fairness in Budget-Constrained Decision Making

The machine learning community and society at large have become increasingly concerned with discrimination and bias in data-driven decision making systems. This has led to a dramatic increase in academic and popular interest in algorithmic fairness. In this work, we focus on fairness in budget-constrained decision making, where the goal is to acquire information (features) one-by-one for each individual to achieve maximum classification performance in a cost-effective way. We provide a framework for choosing a set of stopping criteria that ensures that a probabilistic classifier achieves a single error parity (e.g. equal opportunity) and calibration. Our framework scales efficiently to multiple protected attributes and is not susceptible to intra-group unfairness. Finally, using one synthetic and two public datasets, we confirm the effectiveness of our framework and investigate its limitations.