A multi-head residual connection GCN for EEG emotion recognition

Area of Science:

• Neuroscience
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Electroencephalography (EEG) emotion recognition is vital for intuitive human-computer interaction.
• Existing neural networks struggle to capture deep emotional features from complex EEG data.
• Limitations in conventional methods necessitate advanced techniques for robust emotion classification.

Purpose of the Study:

• To introduce a novel Multi-Head Residual Graph Convolutional Neural Network (MRGCN) for enhanced EEG emotion recognition.
• To leverage complex brain network topology and graph convolution for deeper feature extraction.
• To improve classification accuracy and inter-subject stability in emotion recognition tasks.

Main Methods:

• Decomposition of multi-band differential entropy (DE) features to reveal temporal dynamics of brain activity.
• Integration of short and long-distance brain networks to capture complex topological characteristics.
• Implementation of a residual-based architecture within the graph convolutional network for performance enhancement.

Main Results:

• The MRGCN model achieved high average classification accuracies: 95.8% on the DEAP dataset and 98.9% on the SEED dataset.
• Demonstrated enhanced performance and robustness in EEG-based emotion recognition.
• Visualization of brain network connectivity provided insights into emotional regulation mechanisms.

Conclusions:

• The proposed MRGCN model effectively extracts intricate EEG emotional features, outperforming conventional methods.
• The model's architecture ensures robust and stable emotion classification across different subjects.
• MRGCN offers a promising approach for advancing EEG-based emotion recognition and understanding brain dynamics.