EEG-based Emotion Recognition using Crystal Graph Convolutional Neural Networks with Functional Connectivity and Spatial-Frequency Features

Convolutional neural networks (CNN) have achieved promising performance in EEG-based Emotion Recognition. However, CNNs focus more on learning local features from regular Euclidean data rather than global features. Recently, graph neural networks (GNN) have received much attention in EEG-based emotion recognition because it can handle both local and global features. The proposed functional connectivity and spatial-frequency feature graph neural networks (FCSF-GNN) can handle both local and global features by combining dynamic functional connectivity between different electroencephalogram (EEG) channels and fixed spatial information (physical connectivity). Experimental results show that the proposed method achieves better recognition performance than the state-of-the-art methods in the DEAP dataset. For the tasks of valence and arousal recognition, the proposed method (using 32 channels) achieves the average accuracy of 96.80% and 96.90%, precision of 96.44% and 97.51%, recall of 96.37% and 97.20%, and f1 scores of 96.36% and 97.33%, respectively. Meanwhile, with only about 30%(using 10 channels) of the EEG channels retained, the method achieved an average accuracy of 97.21% and 96.93%, precision of 97.31% and 97.55%, recall of 96.35% and 97.29%, and f1 scores of 96.78% and 97.39%, respectively.