A Feature Selection Method Based On Multi-Objective Genetic Algorithm And Support Vector Machines

Feature selection methods impacts badly on the design and performance of classifiers, it is one of the most important problem in Pattern Recognition, selecting the most discriminative features for classifiers is needed in order to improve the reliability and efficiency of classification algorithms. In this paper, Support Vector Machines are used in feature selection, Support Vector Machines (SVM), one of the new techniques for Pattern Recognition, have been widely used in many application areas. Feature selection and the kernel parameters setting for SVM in the training process are the two important factors that impact the classification accuracy.Feature selection can be viewed as a multi-objective optimization problem, because in the simplest case it involves feature subset size minimization and performance maximization. For the defects such as the bad stability and low classification accuracy of using Standard Genetic Algorithm (SGA) for feature selection and the bad realization of feature selection and SVM parameters optimization synchronously, in this work a feature selection method based on Non-dominated Sorting Genetic Algorithm (NSGA) and SVM is proposed.,NSGA is a good kind of multi-objective genetic algorithms. In the method, firstly, the Wilcoxon-test is used as a coarse gene selection method to remove most of the irrelevant features. Then the value of classification accuracy for single feature is achieved by using leave-one-out cross validation in Support Vector Machines learning process, so the ranking of every feature has got. Finally, NSGA with the embedded SVM, has been used to guide the search towards the more discriminative features and the best number of clusters with the two objectives which are the minimization of the number of features and a validity index error classification that measures the quality of SVM.The proposed strategy is evaluated using two benchmark data sets. Comprehensive experiments demonstrate the feasibility and efficiency of the proposed strategy and the problems exist in using Standard Genetic Algorithm for feature selection have been solved by the method , it is to simultaneously optimize the parameters with improving the generalization performance of SVM and achieve optimal feature subset without degrading the SVM classification accuracy.