Decision Trees with Parametric Enlarged Local Search

Abstract

Lately, decision trees became an important tool in marketing research, e.g. to classify potential customers in order to send individual ads to them. In contrast to many other situations here the classification should be done in such a manner that not the inhomogeneity of subsets is minimised but a cost orientated loss function. Unfortunately, split criteria based on cost orientated loss functions have the disadvantage to be too often indifferent between different splits. This is among others due to the local optimality of the split calculations ([1]) and due to the facts that a further improvement of the loss function could be better achieved if the best split in a “class of indifferent or nearly indifferent splits” is chosen. This paper presents a parametric enlargement of the local search for optimal splits using cost orientated loss functions.