Genetic feature selection for large EEG data with commutation between multiple classifiers

Feature selection represents a key stage in electroencephalogram (EEG) classifications, because these applications involve numerous, high-dimensional samples. In recent literature, a multitude of supervised embedded feature selection procedures has been proposed. Regardless if they are configured as Single Objective (SOO) or Multi-Objective Optimizations (MOO), the embedded methods assess the quality of feature vectors by directly verifying their usefulness on a specific classifier. In this context, defining objective functions independent of the classifier model is almost impossible, meaning that feature selection is inherently disturbed by the incorporated classification method. This paper solves the EEG n-back memory task classification using a new embedded feature selection procedure based on a Genetic Algorithm (GA). The optimization method provides a more robust verification of the competing attributes, by using a switching mechanism which permits assessing the quality of the solutions on multiple classifiers, without significantly increasing the computational time of individuals' evaluation. The approach is exemplified for Random Forests (RF) and Support Vectors Machine (SVM); these two classifiers are interesting for feature selection problems because they offer limited indications regarding the relevance/ irrelevance of the features. This paper also discusses the design of EEG signal preprocessing, meant for enhancing the relevance of the samples for the n-back memory task EEG classification. Lastly, the characteristics of the suggested approach are illustrated by experiments, for two ranking schemes defined in MOO sense.