Classification of Strategic Agents in Gaming and Self-Improvement Systems

Watch a research presentation from the 2022 Symposium on Foundations of Responsible Computing where Carnegie Mellon University researcher Hedyeh Beyhaghi explores the classification challenges of strategic agents who can both game and improve systems. Examine how decision-makers can develop classification rules that minimize false positives while maximizing true positives, particularly in contexts like loan approvals where applicants may take actions to both artificially inflate and legitimately improve their qualifications. Learn about two proposed models - a general discrete model and a linear model - along with their algorithmic implementations, learning capabilities, and computational complexity limitations. Discover efficient algorithms for maximizing true positives without false positives, including extensions to partial-information learning scenarios, while understanding the NP-hardness of certain optimization variations. Gain insights into determining whether linear classifiers can accurately categorize all agents while incentivizing legitimate improvements among those with the capacity to qualify.