A neuroevolutionary approach to feature selection using multiobjective evolutionary algorithms

Abstract

Feature selection plays a central role in predictive analysis where datasets have hundreds or thousands of variables available. It can also reduce the overall training time and the computational costs of the classifiers used. However, feature selection methods can be computationally intensive or dependent of human expertise to analyze data. This study proposes a neuroevolutionary approach which uses multiobjective evolutionary algorithms to optimize neural network parameters in order to find the best network able to identify the most important variables of analyzed data. Classification is done through a Support Vector Machine (SVM) classifier where specific parameters are also optimized. The method is applied to datasets with different number of features and classes.