Predicting outliers

This paper describes a method designed for data mining applications where the main goal is to predict extreme and rare values of a continuous target variable, as well as to understand under which conditions these values occur. Our objective is to induce models that are accurate at predicting these outliers but are also interpretable from the user perspective. We describe a new splitting criterion for regression trees that enables the induction of trees achieving these goals. We evaluate our proposal on several real world problems and contrast the obtained models with standard regression trees. The results of this evaluation show the clear advantage of our proposal in terms of the evaluation statistics that are relevant for these applications.