Abstract
In this paper, we tackle the problem of feature-efficient prediction: classification using a limited number of features per test example. We show that modifying an ensemble classifier such as AdaBoost, by sampling hypotheses from its final weighted predictor, is well-suited for this task. We further consider an extension of this problem, where the costs of examining the various features can differ from one another, and we give an algorithm for this more general setting. We prove the correctness of our algorithms and derive bounds for the number of samples needed for given error rates. We also experimentally verify the effectiveness of our methods.