Spatiotemporal Graph Convolutional Networks for EEG-Based Emotion Recognition

Emotion recognition is a critical task in understanding human affective states and their impact on interactions with products, services, and brands. In this study, we introduce a novel spatiotemporal graph convolutional network (GCN) framework for EEG-based emotion recognition. Unlike conventional CNN and RNN models that struggle with non-Euclidean data structures, our approach leverages the spatial and temporal relationships between EEG channels, captured using advanced GCN techniques. The proposed framework includes spatial and spatiotemporal models, each further divided based on different feature inputs, including Differential Entropy (DE) and Power Spectral Density (PSD). We validate our models on the DEAP dataset, where the spatiotemporal model achieved a valence classification accuracy of 79.7% and an arousal classification accuracy of 68.2%. These results demonstrate that the optimal model configuration significantly enhances emotion classification accuracy, particularly in the recognition of both valence and arousal states. The findings suggest that incorporating GCNs into emotion recognition systems can effectively address the challenges posed by the complex, non-Euclidean structure of EEG data.