SVEM: A Signal Variation Elimination Model for EEG Emotion Recognition

Abstract

Motivated by the non-stationarity characteristics of electroencephalograph (EEG) signals, we propose a signal variation elimination model (SVEM) for emotion recognition. The proposed SVEM enables to capture the topological structures of different EEG channels due to the utilized graph neural network (GNN). Two tricks are proposed to reduce signal variations and improve the model generalization. Firstly, the proposed SVEM is pre-trained by a mask-generation supervised learning where we randomly mask several signal channels in GNN and then generate them. Secondly, the proposed SVEM is fine-tuned by incorporating a domain classifier to reduce the distribution shift between the training and testing sets. To further reduce the subject signal variations of the training set, a subject classifier is incorporated in the fine-tuning process of SVEM. The performance of SVEM is evaluated on the real-world dataset SEED. Experiment results demonstrate that the accuracy of SVEM achieves 87% and 71%, on subject-dependent and subject-independent tasks, respectively.