Feature selection via normative fuzzy information weight with application into tumor classification

Abstract

Feature selection via mutual information has been widely used in data analysing. Mutual information with monotonous is an effective tool to analyse the correlation and redundancy of features. However, the mutual information, which adopted in most of existing feature selection criterions, can’t explain the correlation and redundancy of features in the fuzzy situation well. Therefore, we propose feature selection strategy via normative fuzzy information weight based on fuzzy conditional mutual information in this paper. Firstly, the monotone fuzzy metric structure is defined, and some theoretical properties are proved. Secondly, we put forward the concept of fuzzy independent classification information based on fuzzy conditional mutual information, and propose a feature selection method via fuzzy independent classification information. Thirdly, considering the proportion of new classification information provided by the selected feature in its own information, we introduce the concept of normative fuzzy information weight and propose an improved feature selection method. Finally, the availability of the two proposed methods is tested by comparative experiments, and the improved feature selection method is applied to tumor classification. This work provides an alternative strategy for feature selection in real-world data applications.