Boosting with Diverse Base Classifiers

摘要

We establish a new bound on the generalization error rate of the Boost-by-Majority algorithm. The bound holds when the algorithm is applied to a collection of base classifiers that contains a "diverse" subset of "good" classifiers, in a precisely defined sense. We describe cross-validation experiments that suggest that Boost-by-Majority can be the basis of a practically useful learning method, often improving on the generalization of AdaBoost on large datasets.