Investigation of feature selection algorithms on a cognitive task classification : a comparison study

Abstract

In this study, the effects of feature selection on classification of the electrical signals generated in the brain during numerical and verbal operations are investigated. 18 healthy university/college students were chosen for the experimental study. EEG signals were recorded during silent reading and mental arithmetic operations without using any pen and paper. A total of 60 slides, 30 of which contained reading passages and the rest contained arithmetic operations, were presented in the experiment. EEG signals recorded from 26 channels during the slide show. The recorded EEG signals were analyzed by Hilbert Huang Transform (HHT), and then features were extracted. 312 features were classified by Bayesian Network algorithm without applying feature selection with 92.60% average accuracy. Consistency measures and Correlation based Feature Selection methods were, then, used for feature selection and the numbers of selected features are 8 and 39 on average, respectively. Classification accuracies by using these feature selection algorithms were obtained as 93.98% and 95.58%, respectively. The results showed that feature selection algorithms contribute positively to the classification performance.