Data mining with agent gaming

Abstract

A new type of data mining considers the case where the instances over which induction takes place are intelligent agents who might act strategically to thwart the learner. Instances comprised of humans, companies, or governments all have this capability. One paper calls this adversarial learning and proposes an iterated learning process--much like reinforcement learning--to determine a classifier. The current authors proposed a different approach that uses rational expectation ideas to alter the learner's problem to directly anticipate possible strategic gaming by agents during the induction process. This paper explores differences between solutions produced by these two approaches on a credit dataset and draws some general insights.