Improved Graph Convolutional Neural Networks based on Granger Causality Analysis for EEG Emotion Recognition

Abstract

How to learn the adjacency matrix of graph convolution neural network (GCNN) is one of the key issues in EEG emotion recognition. Most current GCNN models adopt the spatial position between EEG channels to define the adjacency matrix, which have ignored the effective connection characteristics among the EEG signals. Granger causality (GC) analysis can reveal the causal connectivity between EEG channels. In this paper, an improved GCNN method based on the GC analysis is proposed for EEG emotion recognition. Firstly, the causal association matrix is calculated by the GC analysis. Then, a reasonable threshold is used to adaptively converts the causal association matrix into the adjacency matrix, which makes the adjacency matrix more consistent with the cognitive law of the human brain. Finally, the DEAP emotion dataset is used to test the performance of our method. The experimental results demonstrate that the proposed GC-GCNN method achieves better recognition performance than the state-of-the-art methods, in which the average accuracies of 90.11%, 89.48% and 86.35% are respectively obtained for arousal, valence, and arousal-valence classifications.