A study on mutual information-based feature selection in classifiers

Abstract

Multilabel classification is a classification technique in which each sample may be related with more than one class labels. This paper deals with a comparative study of mutual information (MI) technique and other methods in different classifiers. MI is a technique in filter approach type feature selection. It is a fine indicator, which measures the information or data that common between two variables, it audits and evaluates how one of the variables reduces the uncertainty of the other. We consider other two classifiers for the study; they are Naïve Bayesian (NB) and ID3. The experiments were done using MI and compared with the two classifiers for different benchmark data set from UCI repository flag, music, and yeast. The results were verified using evaluation measures. An accurate precision and recall value can be obtained in MI technique rather than using the classifiers NB and ID3.