This paper revisits the problem of optimal learning and decision-making when different misclassification errors incur different penalties. It characterizes when a cost matrix is reasonable and provides a theorem for adjusting the proportion of negative examples in a training set to make optimal cost-sensitive classification decisions using a standard non-cost-sensitive learning method. However, it argues that changing the balance of negative and positive training examples has little effect on classifiers produced by Bayesian and decision tree learning methods. The recommended approach is to learn a classifier from the given training set and then compute optimal decisions using the classifier's probability estimates. The paper also discusses the properties of cost matrices, the importance of measuring costs against a fixed baseline, and the effects of changing base rates on different learning algorithms. It concludes by recommending the use of smoothed probability estimates and empirical threshold adjustment for optimal decision-making in cost-sensitive domains.This paper revisits the problem of optimal learning and decision-making when different misclassification errors incur different penalties. It characterizes when a cost matrix is reasonable and provides a theorem for adjusting the proportion of negative examples in a training set to make optimal cost-sensitive classification decisions using a standard non-cost-sensitive learning method. However, it argues that changing the balance of negative and positive training examples has little effect on classifiers produced by Bayesian and decision tree learning methods. The recommended approach is to learn a classifier from the given training set and then compute optimal decisions using the classifier's probability estimates. The paper also discusses the properties of cost matrices, the importance of measuring costs against a fixed baseline, and the effects of changing base rates on different learning algorithms. It concludes by recommending the use of smoothed probability estimates and empirical threshold adjustment for optimal decision-making in cost-sensitive domains.