Graph Convolutional Neural Network Based on Channel Graph Fusion for EEG Emotion Recognition

To represent the unstructured relationships among EEG channels, graph neural networks are proposed to classify EEG signal. Currently most graph neural networks learn the relationships between EEG channels using a global adjacent matrix of a graph. In fact, a channel is only closely related with a few channels in its neighborhood. Therefore, the local graph structure among EEG channels can also provide useful information for emotion recognition. To solve this issue, we propose an EEG emotion classification model based on channel graph fusion, named DGCN_GF. DGCN_GF can learn dependency relationships among various EEG channels. In DGCN_GF, two kinds of graphs are used to represent channel features. One is the channel local graph, and the other is the channel global graph. We fuse these two kinds of feature representations and use them to recognize EEG emotions. We conduct experiments on the SEED and DREAMER datasets. The experimental results show that the classification accuracy is improved by fusing two different kinds of graph features.