MSLTE: multiple self-supervised learning tasks for enhancing EEG emotion recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) acquisition devices can be unstable, causing information loss in channels or frequency bands.
• Existing EEG emotion recognition models often overlook this instability, leading to overfitting and poor generalization.
• This necessitates robust models capable of handling noisy or incomplete EEG data.

Purpose of the Study:

• To enhance the generalization and reduce overfitting in EEG emotion recognition models.
• To address information loss in EEG data caused by device instability.
• To develop a novel model incorporating self-supervised learning for improved EEG analysis.

Main Methods:

• Introduced channel masking and frequency masking to simulate EEG information loss.
• Developed two self-supervised learning tasks using masked graph autoencoders (GAE) for feature reconstruction.
• Implemented a weight sharing (WS) mechanism between graph decoders for reliable feature reconstruction.
• Utilized an adaptive weight multi-task loss (AWML) strategy combining supervised and self-supervised losses.

Main Results:

• The proposed model achieved higher average emotion classification accuracy across SEED, SEED-V, and DEAP datasets.
• Each module within the model contributed significantly to performance enhancement.
• Demonstrated improved training efficiency, reduced model size, and lower computational complexity compared to state-of-the-art models.
• Exhibited less sensitivity to key parameter variations.

Conclusions:

• Self-supervised learning tasks effectively enhance EEG emotion recognition model generalization and mitigate overfitting.
• The proposed approach offers a more robust and efficient method for EEG-based emotion classification.
• The model's architecture can be adapted for other EEG classification tasks.