Cost-sensitive extensions for global model trees: Application in loan charge-off forecasting

Abstract

Most of regression learning methods aim to reduce various metrics of prediction errors. However, in many real-life applications it is prediction cost, which should be minimized as the under-prediction and over-prediction errors have different consequences. In this paper, we show how to extend the evolutionary algorithm (EA) for global induction of model trees to achieve a cost-sensitive learner. We propose a new fitness function which allows minimization of the average misprediction cost and two specialized memetic operators that search for cost-sensitive regression models in the tree leaves. Experimental validation was performed with bank loan charge-off forecasting data which has asymmetric costs. Results show that Global Model Trees with the proposed extensions are able to effectively induce cost-sensitive model trees with average misprediction cost significantly lower than in popular post-hoc tuning methods.