A Novel and Powerful Dual-Stream Multi-Level Graph Convolution Network for Emotion Recognition.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Emotion recognition is crucial for personalized human-computer interaction.
• Brain activity patterns correlate with cognitive and emotional states.
• Existing models face challenges in capturing complex neural connectivity for emotion recognition.

Purpose of the Study:

• To develop a novel dual-stream multi-level graph convolution network (DMGCN) for enhanced emotion recognition.
• To improve the accuracy and computational efficiency of analyzing brain activity for emotional states.
• To capture hierarchical connectivity patterns in the cerebral cortex for better emotion detection.

Main Methods:

• The proposed DMGCN model integrates a hierarchical dynamic geometric interaction neural network (HDGIL) and a multi-level feature fusion classifier (M2FC).
• HDGIL learns emotion-related representations across multi-level graphs.
• M2FC fuses early and late features from electroencephalogram (EEG) samples for detailed representations.

Main Results:

• The DMGCN model achieved superior classification accuracies across multiple datasets: 98.73% (DEAP-Arousal), 95.97% (DEAP-Valence), 72.74% (DEAP), and 94.89% (SEED).
• The model demonstrated significant improvements over state-of-the-art baselines, with accuracy increases up to 3.17%.
• Experiments validated the effectiveness of individual modules within the DMGCN for emotion recognition tasks.

Conclusions:

• The DMGCN model represents a significant advancement in machine-based emotion recognition using brain activity.
• The network's ability to capture hierarchical neural connectivity and fuse multi-level features enhances understanding of user emotional states.
• The findings highlight the potential of advanced graph neural networks for developing more natural and personalized human-computer interactions.