Classifier-independent feature selection on the basis of divergence criterion

Feature selection aims to choose a feature subset that has the most discriminative information from the original feature set. In practical cases, it is preferable to select a feature subset that is universally effective for any kind of classifier because there is no underlying information about a given dataset. Such a trial is called classifier-independent feature selection. We took notice of Novovicova et al.'s study as a classifier-independent feature selection method. However, the number of features have to be selected beforehand in their method. It is more desirable to determine a feature subset size automatically so as to remove only garbage features. In this study, we propose a divergence criterion on the basis of Novovicova et al.'s method.