| High-dimensional feature spaces are common in applications in various fields of today’s society.However,any such feature space may contain many redundant and irrelevant features,which increases the complexity of data processing.Therefore,it is important to reduce the dimensionality of the feature space while preserving the actual features of the data.Feature selection is one of the most effective tools for eliminating irrelevant and redundant features while retaining relevant features.For this purpose,this paper focuses on two aspects.Traditional information-theoretic-based feature selection methods usually use discrete features,which incur additional discretization overhead when dealing with continuous features.In contrast,fuzzy information-theoretic-based feature selection methods can be applied directly to continuous features.Therefore,two novel fuzzy information-theoreticbased feature selection methods are proposed in this paper.(1)Fuzzy dynamic interactive weighted feature selection method: The existing feature selection methods based on fuzzy information metrics only consider the feature relevance and redundancy,ignoring the dynamic interaction between the candidate features and the selected features.For this reason,a feature selection method based on fuzzy interaction information is proposed.First,the concepts of feature relevance,feature redundancy,and feature interaction are characterized by using fuzzy information-theoretic metrics.Second,a fuzzy interaction weighting factor is defined using fuzzy interaction information,which quantifies the redundancy and interaction among features.Third,a new feature selection algorithm called fuzzy dynamic interaction weighted feature selection(FDIWFS)is designed by combining the fuzzy interaction weight factor with a sequential forward search strategy.To verify the effectiveness of FDIWFS,it is compared with eight state-of-the-art feature selection methods on fifteen publicly available datasets.The results of the comparison experiments show that FDIWFS has better performance in terms of classification accuracy,AUC values and F1 values.(2)Class-specific fuzzy information measure feature selection: Classification-specific feature selection can select the same feature subset for all decision variables,and classspecific feature selection can select different feature subsets for each decision variable.The most existing fuzzy information-theoretic-based feature selection methods are classificationspecific feature selection methods.However,usually a subset of features may not contain all the features required for a single decision variable.In this case,first,we propose a classspecific fuzzy information measure based on the fuzzy information measure.Second,a classspecific feature selection based on the class-specific fuzzy information measure is proposed,called class-specific fuzzy information measure feature selection(CSFIMFS).This method maximizes the relevant information between the selected features and decision variables,while minimizing the redundant information among all features.Finally,we provide a new general framework for methods that aggregate the results produced by each subset of features.We used four classifiers to experiment on twenty-one benchmark datasets,comparing eight classification-specific methods with seven class-specific methods,and found that CSFIMFS showed substantial improvements in classification accuracy,AUC values,and F1 values. |
