End-To-End Selective Decision-Making

Abstract

Artificial intelligence (AI) is increasingly used in high-risk public decision-making, where policymakers face a fundamental trade-off between coverage, the number of decisions made autonomously by AI, and accuracy, the proportion of correct decisions. Higher coverage improves scalability but risks errors, while prioritizing accuracy increases reliance on costly and slow human review. We develop an end-to-end selective decision-making framework that optimizes this coverage-accuracy trade-off in human-AI systems. We formulate selective classification as a mixed-integer linear program and develop a surrogate that learns a classifier maximizing coverage subject to a constraint on false selection risk. By integrating selective classification, end-to-end learning, and conformal inference, our approach provides finite-sample, distribution-free control of false selection risk, rather than relying on heuristic confidence thresholds. We evaluate the method through synthetic experiments and a real-world case study on New York City speed hump requests.

Bio

Michael is an Assistant Professor at Baruch College, CUNY. His research focuses on developing data-driven decision-making methods that utilize decision-aware or end-to-end learning. These methods are essential for solving optimization problems in data-scarce settings, which arise when data collection is expensive, signal-to-noise ratios are low, or systems are highly time inhomogeneous. He has developed data-driven algorithms for problems in urban planning, healthcare, and logistics, with work that has appeared in Operations Research and Management Science. He received his Ph.D. from the Data Sciences and Operations department at the University of Southern California Marshall School of Business and earned his M.S. and B.S. in Operations Research at Columbia University.