EEG emotion recognition based on dynamic temporal causal graph convolutional network

Abstract

Inspired by the connectivity characteristics of brain networks, the dynamic evolution in the connectivity relationships between different brain regions throughout time provides information on emotional representation. However, current electroencephalographic (EEG) emotion recognition methods often overlook local density, global sparsity, and temporal causality that are inherent in the connectivity relationships between brain regions. To address these issues, a dynamic temporal–causal graph convolutional network (DTC-GCN) for EEG emotion recognition is proposed herein. The DTC-GCN learns the spatial topology and temporal–causal relationships between EEG channels over a period of time. It takes time-series graphs as the input and is implemented in two stages. In the first stage, the sparse connected dynamic graph convolutional network is used to dynamically learn a locally dense and global sparse brain network. In the second stage, the temporal–causal module is used to construct causal connectivity among EEG channels across different segments. The effectiveness of the proposed model is evaluated by conducting extensive experiments on two publicly available datasets, DEAP and SEED. On the DEAP dataset, the average accuracies of arousal and valence are 95.08% and 94.31%, respectively. On the SEED dataset, the average accuracy is 98.48%. Results indicate that the DTC-GCN outperforms existing state-of-the-art methods. By analyzing the parameters of the DTC-GCN and conducting an interpretability study, we reveal the overall connectivity pattern between EEG channels and the causal relationships between segments within short time intervals.