USING C4.5 AS VARIABLE SELECTION CRITERION IN CLASSIFICATION TASKS

Variable selection is a difficult and important problem in machine learning. For classification tasks, it can lead to in-creased accuracy or to reduced computational costs. In this paper we present an experimental study that shows how a very simple heuristic, namely using C4.5 for variable selec-tion, can maintain classification accuracy in many bench-mark problems while significantly reducing running times. In addition, we construct an ensemble that combines clas-sifiers using the variables selected by C4.5 with classifiers that use the full variable set. Experimental results show that by using the selected variable set with C4.5, the clas-sification accuracy is similar to that obtained by using the full variable set. This suggests that using C4.5 is a good approach for variable selection in classification tasks