This paper proposes cost sensitive criteria for constructing classification rules by supervised learning methods. Reinterpreting established loss functions and considering those introduced by Buja, Stuetzle, et al. (2005) and Hand (2009), we identify criteria reflecting different degrees of information about misclassification costs. To adapt classification methodology to practical cost considerations, we suggest the use of these criteria for different model selection approaches in supervised learning. In addition, we investigate the effects of cost sensitive adaptations in CART and boosting and conclude that adaptations are more promising in the selection rather than in the estimation step.
unequal misclassification costs proper scoring rules AUC boosting CART model selection pruning early stopping