Multi-Domain Based Dynamic Graph Representation Learning for EEG Emotion Recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Graph neural networks (GNNs) show promise for electroencephalogram (EEG) emotion recognition due to efficient graph-structured data processing.
• Challenges exist in representing EEG as graph data due to dynamic functional connectivity and nonlinear relationships between brain regions.

Purpose of the Study:

• To propose a novel Multi-Domain based Graph Representation Learning (MD²GRL) framework for modeling EEG signals as graph data.
• To enhance emotion recognition accuracy from EEG signals by addressing challenges in graph representation.

Main Methods:

• MD²GRL utilizes gated recurrent units (GRU) and power spectral density (PSD) for node feature construction in two subgraphs.
• A self-attention mechanism learns node similarity, fused with spatial information to create an adjacency matrix.
• A learnable soft thresholding operator sparsifies the adjacency matrix, and a dual-branch GNN with spatial asymmetry is used for classification.

Main Results:

• The proposed method achieved state-of-the-art (SOTA) classification performance on SEED and DEAP datasets for both subject-dependent and independent tasks.
• Visualization analysis identified significant EEG channel connections related to emotion, effectively suppressing irrelevant noise.
• The learned graph structures align with neuroscience findings, indicating the model's ability to capture neural underpinnings of emotion.

Conclusions:

• The MD²GRL framework provides an effective approach for representing EEG signals as graphs, overcoming previous limitations.
• The model demonstrates high accuracy in emotion classification, offering insights into brain connectivity patterns associated with emotional states.
• This approach holds significant potential for advancing research into the neural basis of emotion using EEG data.