Multi-source domain adaptation for EEG emotion recognition based on inter-domain sample hybridization.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) is crucial for emotion recognition but suffers from inter-subject and inter-session variability.
• Existing domain adaptation (DA) methods align distributions but can confuse emotion labels.
• This work addresses the challenge of robust EEG-based emotion recognition across different individuals and recording sessions.

Purpose of the Study:

• To improve cross-subject and cross-session EEG emotion recognition by enhancing domain adaptation.
• To address the label confusion issue in current DA methods by promoting conditional distribution alignment.
• To develop a robust model for EEG emotion recognition that generalizes across subjects and sessions.

Main Methods:

• Introduced a multi-source domain adaptation common-branch network for EEG emotion recognition.
• Proposed a novel sample hybridization method to incorporate target domain information without increasing data size.
• Validated the model using cross-subject and cross-session experiments on SEED and SEED-IV datasets.

Main Results:

• Achieved 90.27% average accuracy in cross-subject emotion recognition on the SEED dataset.
• Reached 73.21% accuracy on the SEED-IV dataset for cross-subject recognition.
• Obtained average accuracies of 94.16% and 75.05% in cross-session experiments on SEED and SEED-IV, respectively.

Conclusions:

• The proposed method effectively transfers emotion-related knowledge from source to target domains.
• The combination of multi-source DA and sample hybridization enhances generalization ability for EEG emotion recognition.
• The approach successfully achieves accurate emotion recognition for unlabeled subjects, overcoming inter-subject and inter-session limitations.