Path Adversarial Dual-Branch Network for EEG Emotion Recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG)-based emotion recognition faces challenges with cross-subject domain shift.
• Insufficient complementary fusion of time-frequency information hinders accurate emotion classification.
• Existing models struggle with aligning feature distributions across different domains and modalities.

Purpose of the Study:

• To propose a novel multi-task adversarial network, the Path Adversarial Dual-Branch Network (PADB-Net), for robust EEG-based emotion recognition.
• To address cross-subject domain shift and improve the fusion of time-frequency information in EEG signals.
• To align feature distributions across time and frequency domains and between source and target domains.

Main Methods:

• A dual-branch parallel architecture processes raw EEG waveforms (time domain) and differential entropy features (frequency domain).
• Lightweight depthwise separable convolutions and channel attention are employed for discriminative feature extraction.
• A path adversarial module and a domain adversarial module are introduced to align feature distributions within a unified framework.

Main Results:

• PADB-Net significantly outperforms single-adversarial and non-adversarial baselines in accuracy, AUC, F1-score, sensitivity, and specificity.
• On the HybridBCI dataset, PADB-Net achieved 77.80% accuracy, 84.50% AUC, and 79.40% F1-score with minimal parameters.
• The model demonstrated strong cross-dataset generalizability on the SEED dataset, achieving high F1-scores for negative, neutral, and positive emotions.

Conclusions:

• The proposed PADB-Net effectively mitigates cross-subject domain shift and enhances time-frequency information fusion in EEG emotion recognition.
• The synergistic gain of the dual-adversarial mechanism is verified, leading to significant performance improvements.
• PADB-Net shows promising results and strong generalizability for real-world emotion recognition applications.