Feature Selection Using Mutual Information: An Experimental Study

In real-world application, data is often represented by hundreds or thousands of features. Most of them, however, are redundant or irrelevant;, and their existence may straightly lead to poor performance of learning algorithms. Hence, it; is a compelling requisition for their practical application,,, to choose most salient features. Currently, a large number of feature selection methods using various strategies have been proposed. Among these methods, the mutual information ones have recently gained much more popularity. In this paper, a general criterion function for feature selector using mutual information is firstly introduced. This function can bring up-to-date selectors based on mutual information together under an unifying scheme. Then an experimental comparative study of eight typical filter mutual information based feature selection algorithms on thirty-three datasets is presented. We evaluate them from four essential aspects, and the experimental results show that none of these methods outperforms others significantly. Even so, the conditional mutual information feature selection algorithm dominates other methods on the whole, if training time is not a matter.