EEG Emotion Recognition based on Hierarchy Graph Convolution Network

Emotion recognition has become a research focus in the field of human-computer interaction (HCI). As an excellent physiological signal, electroencephalographic (EEG) is considered to be a favorable tool for emotion recognition. Most traditional methods focus on extracting features in time domain and frequency domain but the adjacent information and asymmetric information from adjacent and asymmetric channels are often ignored. Although several graph neural network (GNN) models are utilized to learn EEG features, most of the emotion recognition studies of GNN ignore the information existing between adjacent electrodes. In this paper, we propose an EEG emotion recognition method based on hierarchy graph convolution network (HGCN) named ERHGCN. Firstly, six different features including power spectral density (PSD), differential entropy (DE), differential asymmetry (DASM), rational asymmetry (RASM), asymmetry (ASM) and differential caudality (DCAU) from five frequency bands are extracted. Secondly, to improve graph convolution network (GCN) shortcoming of only extracting time and frequency features, HGCN is applied to extract deeper spatial feature by treating the longitudinal and transverse adjacent electrode pairs in different ways. Finally, six extracted features are fed into the HGCN model, then all features are integrated by two full connection layers. We conducted extensive experiments on DEAP dataset and experimental results show that the proposed method can obtain 90.56% and 88.79% recognition accuracies for valence and arousal classification tasks.