Semi-supervised EEG emotion recognition model based on enhanced graph fusion and GCN

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) based emotion recognition faces challenges due to signal instability and emotional complexity.
• Existing Graph Convolutional Network (GCN) models struggle with single-feature representation and graph noise.
• Need for advanced strategies like feature/similarity fusion and noise reduction in EEG emotion recognition.

Purpose of the Study:

• To develop a semi-supervised EEG emotion recognition model leveraging both labeled and unlabeled data.
• To enhance GCN performance by addressing limitations of single-feature representation and graph noise.
• To introduce novel techniques for graph fusion, network enhancement, and feature fusion.

Main Methods:

• Extracted and compacted EEG features using Principal Component Analysis (PCA).
• Constructed Sample-by-sample Similarity Matrices (SSM) and fused graphs using Similarity Network Fusion (SNF).
• Applied Network Enhancement (NE) for noise reduction and utilized GCN for feature propagation on fused data.

Main Results:

• Achieved significant classification accuracy improvements: 1.52% on SEED and 13.14% on SEED-IV with minimal labeled data.
• Reduced training time by 46.75s (SEED) and 22.55s (SEED-IV) compared to state-of-the-art methods.
• Demonstrated that graph fusion, network enhancement, and feature fusion collectively enhance model performance.

Conclusions:

• The proposed semi-supervised model effectively integrates graph fusion, network enhancement, and feature fusion for superior EEG emotion recognition.
• The combined approach significantly boosts accuracy and reduces computational cost, outperforming existing methods.
• Key hyperparameters are identifiable and easily adjustable for optimal performance, simplifying model implementation.