Dual pseudo-labeling based adversarial domain adaptation for EEG-based emotion recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Unsupervised domain adaptation (UDA) is crucial for cross-subject emotion recognition using electroencephalography (EEG).
• Existing UDA methods often neglect target domain class information, leading to coarse alignment and misclassification.
• There is a need for more refined UDA techniques to improve the generalization of EEG-based emotion recognition.

Purpose of the Study:

• To propose a novel adversarial training-based domain adaptation framework for cross-subject emotion recognition.
• To enhance intra-class correlation and inter-class discriminability using emotion class prototypes and soft pseudo-labels.
• To improve the robustness and quality of pseudo-labels through a dual pseudo-labeling strategy for fine-grained domain alignment.

Main Methods:

• Leveraging emotion class prototypes to strengthen feature distribution correlation between source and target domains.
• Utilizing soft pseudo-labels from prototype clustering to boost within-domain class separability.
• Implementing a dual pseudo-labeling strategy to enhance pseudo-label quality and adversarial training for fine-grained distribution alignment.

Main Results:

• The proposed method demonstrates significant effectiveness in cross-subject and cross-session emotion recognition evaluations.
• Experimental results show advantages over several state-of-the-art unsupervised domain adaptation approaches.
• The dual pseudo-labeling strategy provides additional supervision, refining domain adaptation and improving model generalization.

Conclusions:

• The novel adversarial training framework with prototype-based adaptation and dual pseudo-labeling effectively addresses limitations of existing UDA methods.
• The approach achieves a more fine-grained domain alignment, leading to improved accuracy in EEG-based emotion recognition.
• This study significantly enhances the generalization capability of cross-subject emotion recognition models.