Strategic Classification

Strategic classification studies how people alter features or behavior in response to classifiers, and how to design decision systems that account for those responses. Approaches model agents’ costs and best responses, then train classifiers for post-manipulation outcomes using robust optimization, causal models, implicit differentiation, and online or PAC learning. Work also examines policy and interaction levers such as disclosures, audits, abstention, multi-stage decisions, and fairness constraints, with evaluation on synthetic and real-world tabular data. A central tension is curbing harmful gaming while preserving genuine improvement, fairness, welfare, and useful information when behavior and costs may be uncertain or interdependent.