EEG emotion recognition based on Ordinary Differential Equation Graph Convolutional Networks and Dynamic Time Wrapping

Abstract

Graph Convolutional Network (GCN) has been extensively utilized to extract relations among electroencephalography (EEG) electrode channels for its strong ability to handle non-Euclidean data. However, GCN still has some issues when it comes to extracting features from EEG signals: (1) GCN with more layers may experience over-smoothing, restricting its ability to mine longer dependency relations. (2) At the moment, most GCNs used to process EEG signals construct adjacency matrices by Euclidean distance, only considering the correlations on the feature domain while ignoring changes of signals over the entire time window. To address the issues above, we introduce an Ordinary Differential Equation (ODE) based GCN, which can perfectly eliminate the over-smoothing problem of the traditional GCN. Besides, we also propose a method based on Dynamic Time Wrapping (DTW) algorithm to construct an adjacency matrix in the time domain. To handle adjacency matrices calculated by Euclidean distance and DTW distance respectively, we apply a temporal–spatial model composed of two parallel modules each containing an ODE-based GCN and Long short-term memory neural networks (LSTM) network in turn. We conducted experiments on three public datasets. The results show that our methods have achieved an improvement of 2.19%/2.77%/2.13%/2.01% on Arousal/Valence/Dominance/Liking on DEAP dataset, 1.43% on SEED dataset and 3.06%/3.27% on Arousal/Valence on DREAMER dataset compared with state-of-the-art (SOTA) baseline methods. It demonstrates that our method can effectively approve the performance to handle the relations between EEG channels. The premise of the ODE-based GCN is that signal changes of all EEG channels should be continuous rather than abrupt. We believe that it conforms to the EEG mode, as it is activated by the same emotion stimulation while being collected.