Hierarchical Model for Rank Discrimination Measures
Résumé
In this paper we focus on rank discrimination measures, i.e., functions able to quantify the discrimination power of an attribute w.r.t. the class, taking into account the monotonicity of the class w.r.t. the attribute. These measures are used in decision tree induction in order to enforce a local form of monotonicity of the class w.r.t. the splitting attribute and are characterized by a noticeable robustness to non-monotone noise present in the data. More precisely, here we present a hierarchical model in order to single out which properties a function must satisfy to be a rank discrimination measure, providing in this way a framework for the construction of new measures.