Bi-Hemispheric Adversarial Domain Adaptation Neural Network for EEG-Based Emotion Recognition

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG)-based emotion recognition faces challenges due to individual differences and signal non-stationarity.
• Existing adversarial domain adaptation methods often overlook the multimodal structure of EEG data and hemispheric asymmetries.
• Global distribution alignment using binary discriminators may not fully capture complex emotional EEG characteristics.

Purpose of the Study:

• To propose a Bi-Hemispheric Adversarial Domain Adaptation Neural Network (BiHADA) for enhanced EEG-based emotion recognition.
• To address limitations of existing methods by incorporating label structure and modeling hemispheric differences.
• To improve the transferability and discriminability of EEG emotional features across domains.

Main Methods:

• Developed a multimodal discriminator integrating source-domain label structure for improved domain alignment.
• Implemented dual adversarial domain adaptation branches to separately model left and right hemispheric processing.
• Introduced discriminator-derived perplexity for adaptive weighting of hemisphere classifiers based on alignment quality.

Main Results:

• BiHADA achieved 86.82% accuracy in cross-subject tasks and 92.71% in cross-session tasks on the SEED dataset.
• Demonstrated significant improvements in the transferability and discriminability of EEG emotional features.
• The method effectively handles different domain adaptation scenarios in EEG emotion recognition.

Conclusions:

• BiHADA enhances EEG emotion recognition by combining class-structure-guided domain alignment and hemispheric asymmetry modeling.
• Incorporating source-domain label structure and hemisphere-specific adaptation boosts cross-domain performance.
• The proposed approach offers a more robust framework for understanding and recognizing emotions from EEG signals.