Feature selection method for dynamic feature interaction in fuzzy neighborhood system

In order to solve the problem that most feature selection methods based on fuzzy neighborhood information system only monotonically interact with features according to the order of features when selecting features, and that most feature measurement functions only construct metric functions from the perspective of algebraic view or information view, a feature selection method based on dynamic feature interaction in fuzzy neighborhood is proposed. Firstly, fuzzy neighborhood mutual information is introduced to calculate the feature correlation degree and reorganize the feature order according to the feature correlation degree. Secondly, the process of dynamic interaction between features is analyzed, and the degree of redundancy and dynamic interaction between features is calculated according to the order of features through fuzzy neighborhood mutual information and fuzzy neighborhood conditional mutual information. Finally, in order to improve the defect of single perspective in the construction of most feature metric functions, a hybrid mutual information metric function relying on fuzzy neighborhoods from multiple perspectives is proposed. Experimental results show that the proposed algorithm eliminates redundant features and improves the accuracy of data classification while comparing with seven existing feature reduction algorithms on eight public datasets.