Algorithmic decision making and the cost of fairness. Sam Corbett-Davies, Emma Pierson, Avi Feller, Sharad Goel, and Aziz Huq. 2017. Algorithmic decision making and the cost of fairness. In Proceedings of KDD '17, Halifax, NS, Canada, August 13-17, 2017, 10 pages. DOI: 10.1145/3097983.3098095 Algorithms are now used to decide whether defendants awaiting trial are too dangerous to be released. Black defendants are more likely to be incorrectly classified as high risk than white defendants. To reduce racial disparities, several techniques have been proposed to achieve algorithmic fairness. The authors reformulate algorithmic fairness as constrained optimization: the objective is to maximize public safety while satisfying formal fairness constraints. They show that for several past definitions of fairness, the optimal algorithms require detaining defendants above race-specific risk thresholds. They further show that the optimal unconstrained algorithm requires applying a single, uniform threshold to all defendants. This safety-maximizing rule satisfies one important understanding of equality: that all individuals are held to the same standard, irrespective of race. Because the optimal constrained and unconstrained algorithms generally differ, there is tension between improving public safety and satisfying prevailing notions of algorithmic fairness. By examining data from Broward County, they show that this trade-off can be large in practice. They focus on algorithms for pretrial release decisions, but the principles apply to other domains and to human decision makers. The authors define three popular definitions of algorithmic fairness: statistical parity, conditional statistical parity, and predictive equality. They show that the optimal algorithms for these definitions are simple, deterministic threshold rules based on p_{Y|X}. For statistical parity and predictive equality, the optimal algorithms detain defendants when p_{Y|X} exceeds a group-specific threshold. For conditional statistical parity, the thresholds depend on both group membership and "legitimate" factors. The unconstrained algorithm applies a single, uniform threshold to all individuals. The authors prove that threshold rules are optimal under these fairness criteria. They also show that the optimal constrained algorithms differ from the optimal unconstrained algorithm, meaning fairness has a cost. The authors analyze data from Broward County to show that satisfying common fairness definitions can result in detaining low-risk defendants and reducing public safety. They find that enforcing statistical parity leads to a 9% increase in violent recidivism among released defendants. They also show that a single-threshold rule that maximizes public safety generally violates all three fairness definitions. For example, in the Broward County data, optimally detaining 30% of defendants with a single-threshold rule means that 40% of black defendants are detained, compared to 18% of white defendants, violating statistical parity. And among defendants who ultimately do notAlgorithmic decision making and the cost of fairness. Sam Corbett-Davies, Emma Pierson, Avi Feller, Sharad Goel, and Aziz Huq. 2017. Algorithmic decision making and the cost of fairness. In Proceedings of KDD '17, Halifax, NS, Canada, August 13-17, 2017, 10 pages. DOI: 10.1145/3097983.3098095 Algorithms are now used to decide whether defendants awaiting trial are too dangerous to be released. Black defendants are more likely to be incorrectly classified as high risk than white defendants. To reduce racial disparities, several techniques have been proposed to achieve algorithmic fairness. The authors reformulate algorithmic fairness as constrained optimization: the objective is to maximize public safety while satisfying formal fairness constraints. They show that for several past definitions of fairness, the optimal algorithms require detaining defendants above race-specific risk thresholds. They further show that the optimal unconstrained algorithm requires applying a single, uniform threshold to all defendants. This safety-maximizing rule satisfies one important understanding of equality: that all individuals are held to the same standard, irrespective of race. Because the optimal constrained and unconstrained algorithms generally differ, there is tension between improving public safety and satisfying prevailing notions of algorithmic fairness. By examining data from Broward County, they show that this trade-off can be large in practice. They focus on algorithms for pretrial release decisions, but the principles apply to other domains and to human decision makers. The authors define three popular definitions of algorithmic fairness: statistical parity, conditional statistical parity, and predictive equality. They show that the optimal algorithms for these definitions are simple, deterministic threshold rules based on p_{Y|X}. For statistical parity and predictive equality, the optimal algorithms detain defendants when p_{Y|X} exceeds a group-specific threshold. For conditional statistical parity, the thresholds depend on both group membership and "legitimate" factors. The unconstrained algorithm applies a single, uniform threshold to all individuals. The authors prove that threshold rules are optimal under these fairness criteria. They also show that the optimal constrained algorithms differ from the optimal unconstrained algorithm, meaning fairness has a cost. The authors analyze data from Broward County to show that satisfying common fairness definitions can result in detaining low-risk defendants and reducing public safety. They find that enforcing statistical parity leads to a 9% increase in violent recidivism among released defendants. They also show that a single-threshold rule that maximizes public safety generally violates all three fairness definitions. For example, in the Broward County data, optimally detaining 30% of defendants with a single-threshold rule means that 40% of black defendants are detained, compared to 18% of white defendants, violating statistical parity. And among defendants who ultimately do not