GCNFormNet: branched graph-transformer architecture for EEG-based emotion recognition

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Existing electroencephalography (EEG)-based emotion recognition methods often require complex preprocessing, hindering the evaluation of core architectural capabilities.
• There is a need for advanced architectures that can effectively model both spatial and temporal dependencies within EEG signals without extensive preprocessing.

Purpose of the Study:

• To introduce GCNFormNet, a hybrid deep learning architecture designed for robust EEG-based emotion recognition.
• To evaluate the performance and interpretability of GCNFormNet across multiple benchmark EEG datasets.

Main Methods:

• GCNFormNet employs a dual-branched design: Graph Convolutional Network (GCN) layers for spatial relationships and Transformer blocks (Performer-based self-attention) for temporal dynamics.
• A dynamically generated adjacency matrix is used in GCN layers, and DynamicTanh (DyT) replaces traditional layer normalization in Transformer blocks.
• The architecture was evaluated using the EEGain framework on SEED, SEED-IV, DEAP, and DREAMER datasets.

Main Results:

• GCNFormNet achieved competitive performance, notably the highest accuracy (0.46) on the SEED-IV dataset, confirmed by statistical significance.
• Interpretability analysis of learned adjacency matrices revealed neurophysiologically relevant brain connectivity patterns, such as hemispheric asymmetry and prefrontal dominance, without prior anatomical information.
• Ablation and sensitivity analyses confirmed the synergistic effect of GCN and Transformer components on most datasets, though a dataset-specific dependency was noted for DEAP.

Conclusions:

• GCNFormNet offers a promising approach to EEG-based emotion recognition by effectively modeling spatio-temporal dynamics and reducing reliance on complex preprocessing.
• The model's ability to uncover meaningful neurophysiological patterns highlights its potential for both emotion recognition and brain connectivity research.
• Further investigation into dataset-specific optimizations may be warranted for diverse EEG applications.