Multi-source domain transfer network based on subdomain adaptation and minimum class confusion for EEG emotion...

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Electroencephalogram (EEG) signals are crucial for objective brain state reflection in emotion recognition research.
• Cross-session and cross-subject variability in EEG data present significant challenges for practical emotion recognition systems.

Purpose of the Study:

• To propose a novel multi-source domain transfer method, MS-SAMCC, to overcome EEG variability issues in emotion recognition.
• To enhance the robustness and generalizability of EEG-based emotion recognition models.

Main Methods:

• Implemented a mix-up data augmentation technique for generating synthetic EEG samples.
• Developed a Minimum Class Confusion Subdomain Adaptation (MCCSA) method for global and subdomain alignment between source and target domains.
• Utilized Minimum Class Confusion (MCC) as a regularizer within the subdomain adaptation module.

Main Results:

• Achieved high cross-subject accuracies: 87.14% (SEED), 63.24% (SEED IV), and 42.07% (FACED).
• Obtained strong cross-session accuracies: 94.20% (SEED) and 71.66% (SEED IV).
• Demonstrated the effectiveness of the MS-SAMCC approach in diverse EEG emotion recognition tasks.

Conclusions:

• The proposed MS-SAMCC method effectively mitigates cross-session and cross-subject variations in EEG signals.
• This approach significantly improves the performance of emotion recognition systems, paving the way for practical applications.
• The study highlights the potential of domain adaptation techniques in advancing brain-computer interfaces.