Pruning Improves Heuristic Search for Cost-Sensitive Learning

Abstract

This paper addresses cost-sensitive classification in the setting where there are costs for measuring each attribute as well as costs for misclassification errors. We show how to formulate this as a Markov decision problem in which the transition model is learned from the training data. Specifically, we assume a set of training examples in which all attributes (and the true class) have been measured. We describe a learning algorithm based on the AO heuristic search procedure that searches for the classification policy with minimum expected cost. Although we provide a good admissible heuristic for AO , the search space still explodes for realistic-sized problems. To tame this explosion, we introduce a pruning heuristic based on the principle that if the values of two policies are statistically indistinguishable (on the training data), then we can prune one of the policies from the AO search space. Experiments with realistic and synthetic data demonstrate that this pruning heuristic can substantially reduce the memory needed for AO search without significantly affecting the quality of the learned policy. Hence, statistical pruning expands the range of cost-sensitive learning problems for which AO is feasible.