AGCN-SAT: Adaptive Graph Convolutional Network with Spatial Attention and Transformer for EEG Emotion Recognition

EEG is the most commonly used input signal in emotion recognition tasks. Since EEG has topological structure, using graph convolutional network is a more appropriate approach. However, how to capture the dynamic changes of dependency information between EEG channels in different emotional states is still a challenge. In addition, graph convolution only captures the functional connectivity relationships between local channels, and misses many useful global spatial information. To solve the above problems, this paper proposes an Adaptive Graph Convolutional Network with Spatial Attention and Transformer (AGCN-SAT) for EEG emotion recognition. The model constructs an adaptive learning adjacency matrix to facilitate better mining of local spatial features by the graph convolutional network. Meanwhile, the Transformer module is applied to extract global spatial features and concatenate them with local spatial features to form the complete spatial features. Also, spatial attention mechanism is introduced to make model focus more attention on EEG channels related to target emotion, so as to obtain more discriminative features. We conduct experiments on SEED dataset, and the experimental results show that the proposed method achieves excellent performance.