Strategic Classification

Hardt et al. (2015) address classifier manipulation by strategic actors, modeling the problem as a sequential game between classifier designers and individuals seeking favorable classification who may alter attributes to game the system. For natural cost function classes, they developed computationally efficient algorithms achieving near-optimal performance, though designing strategy-proof classifiers for general cost functions with inverse-polynomial approximation is NP-hard. This work fundamentally addresses the tension between classification accuracy and robustness when subjects have manipulation incentives.