Approximation algorithms for reducing classification cost in ensembles of classifiers

We develop approximation algorithms for reducing expected classification cost, when there is a cost associated with obtaining the value of each attribute, and we are doing classification based on an ensemble of linear threshold classifiers. We focus on the stochastic setting where attribute values are independent, and their distributions are given. We review related work based on reductions to Stochastic Submodular Set Cover. We prove approximation bounds for determining the majority prediction of the classifiers in the ensemble, and for detemrining the prediction of at least one classifier.