Correlation-based neural gas for visualizing correlations between EEG features

Abstract

Feature selection is an important issue in an automated data analysis. Unfortunately the majority of feature selection methods does not consider inner relationships between features. Furthermore existing methods are based on a prior knowledge of a data classification. Among many methods for displaying data structure there is an interest in self organizing maps and its modifications. Neural gas network has shown surprisingly good results when capturing the inner structure of data. Therefore we propose its modification (correlation - based neural gas) and we use this network to visualize correlations between features. We discuss the possibility to use this additional information for fully automated unsupervised feature selection where no classification is available. The algorithm is tested on the EEG data acquired during the mental rotation task. © 2013 Springer-Verlag Berlin Heidelberg.