We present a new algorithm for the contextual bandit learning problem, where the learner repeatedly takes an action in response to the observed context, observing the reward only for that action. Our method assumes access to an oracle for solving cost-sensitive classification problems