SCC-MPGCN: self-attention coherence clustering based on multi-pooling graph convolutional network for EEG emotion recognition

The emotion recognition with electroencephalography (EEG) has been widely studied using the deep learning methods, but the topology of EEG channels is rarely exploited completely. In this paper, we propose a self-attention coherence clustering based on multi-pooling graph convolutional network (SCC-MPGCN) model for EEG emotion recognition. The adjacency matrix is constructed based on phase-locking value to describe the intrinsic relationship between different EEG electrodes as graph signals. The graph Laplacian matrix is obtained from the adjacency matrix and then is fed into the graph convolutional layers to learn the generalized features. Moreover, we propose a novel graph coarsening method called SCC, using the coherence to cluster the nodes. The benefits are that the dimensionality of adjacency matrix can be reduced and the global information can be achieved from the raw data. Meanwhile, a MPGCN block is introduced to learn the generalized features of emotional states. The fully-connected layer and a softmax layer are adopted to derive the final classification results. We carry out the extensive experiments on DEAP dataset and the results show that the proposed method has better classification results than the state-of-the-art methods with the ten-fold cross-validation. And the model achieves the emotion recognition performance with a mean accuracy of 96.37%, 97.02%, 96.72% on valence, arousal, and dominance dimension, respectively.