Incentive-Aware PAC Learning

We study PAC learning in the presence of strategic manipulation, where data points may modify their features certain predefined ways order to receive a better outcome. show that vanilla ERM principle fails achieve any nontrivial guarantee this context. Instead, we propose an incentive-aware version which has asymptotically optimal sample complexity. then focus our attention on incentive-compatible classifiers, provably prevent kind manipulation. give complexity bound is, curiously, independent hypothesis class, for restricted classifiers. This suggests incentive compatibility alone can act as effective means regularization. further it is without loss generality consider only classifiers when opportunities manipulation satisfy transitivity condition. As consequence, such cases, hypothesis-class-independent applies even compatibility. Our results set foundations learning.