Design of Feature Selection Algorithm Based on Improved FCBF

Feature selection is an indispensable step in data mining, and its core is to analyze and quantify the correlation and redundancy between features and classes. Firstly, in the FCBF feature selection method, we use the symmetric uncertainty measurement criterion to evaluate the correlation of each feature and class label, sort each correlation measure from small to large, select the appropriate threshold, and eliminate the features below the threshold; secondly, use the approximate Markov blanket principle to screen irrelevant features and redundant features, and finally select the optimal feature subset. In this feature selection method, the complementarity between features is rarely taken into account, and it is easy to have some complementary features which are helpful to classification in the eliminated redundant features. Therefore, a feature selection method based on correlation, redundancy and complementarity is proposed, which enters the complementarity measurement criterion on the basis of FCBF algorithm to select complementary features and improve the accuracy of feature selection.