Understanding Power of Graph Convolutional Neural Network on Discriminating Human EEG Signal

Electroencephalogram (EEG) based biometric is an emerging field in security systems, which provides a higher reliability compared to the conventional identification methods. EEG devices capture the temporal brain waves of individuals that are unique and cannot be reproduced. The individual variability of EEG signals can be derived from its functional connection, which is best represented by a graph. Although a Graph Convolutional Neural Network (GCNN) has improved the classification of graph data in recent years, GCNN is still not well addressed for the EEG biometric identification system. We introduce a novel GCNN model that overcomes the lack of performance on small graph data. Furthermore, we investigate the performance of recent GCNN benchmarks considering both model loss and complexity. The proposed GCNN results in more that 95 % accuracy with considerably low computational cost for 3 databases recorded in 8 different human states. Our experiment shows that adding an extra graph convolution or pooling layer does not necessarily result in better performance.