MS-MDA: Multisource Marginal Distribution Adaptation for Cross-Subject and Cross-Session EEG Emotion Recognition

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) based emotion recognition is vital for psychiatric disorder diagnosis and rehabilitation.
• Practical application of EEG emotion recognition is hindered by significant inter-subject and inter-session variability.
• Existing domain adaptation (DA) methods often oversimplify by treating diverse EEG data as a single source domain, complicating adaptation.

Purpose of the Study:

• To propose a novel Multi-Source Marginal Distribution Adaptation (MS-MDA) framework for robust EEG-based emotion recognition.
• To address the limitations of existing DA approaches in handling multiple, heterogeneous EEG data sources.
• To improve the accuracy and reliability of emotion recognition across different subjects and recording sessions.

Main Methods:

• Developed MS-MDA to leverage both domain-invariant and domain-specific features from multiple EEG data sources.
• Assumed shared low-level features across EEG data while employing independent branches for one-to-one DA and domain-specific feature extraction.
• Utilized a multi-branch inference strategy for final emotion recognition.

Main Results:

• MS-MDA demonstrated superior performance compared to existing methods in cross-session and cross-subject transfer scenarios.
• Evaluated on the SEED and SEED-IV datasets for recognizing three and four distinct emotions, respectively.
• Achieved state-of-the-art results, outperforming comparison models in challenging transfer learning settings.

Conclusions:

• MS-MDA effectively mitigates the impact of inter-subject and inter-session variability in EEG emotion recognition.
• The proposed approach offers a more sophisticated and effective strategy for domain adaptation in EEG analysis.
• This work provides a promising direction for enhancing the practical utility of EEG-based emotion recognition systems.