Dynamic Domain Adaptation for Class-Aware Cross-Subject and Cross-Session EEG Emotion Recognition

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) based emotion recognition systems require generalizable models for real-world applications.
• Existing domain adaptation methods address global EEG data discrepancies but neglect category-specific local divergences.
• This oversight diminishes the discriminative power of emotion-invariant features.

Purpose of the Study:

• To propose a novel Dynamic Domain Adaptation (DDA) algorithm for improving EEG emotion recognition.
• To enhance feature generalizability and discriminative ability by aligning both global and local EEG data distributions.
• To develop a dynamic training strategy that addresses the lack of target domain emotion labels.

Main Methods:

• The Dynamic Domain Adaptation (DDA) algorithm minimizes global domain discrepancy and local subdomain discrepancy.
• A dynamic training strategy is employed, starting with global alignment and progressing to local alignment.
• Unsupervised and semi-supervised versions of DDA are implemented for diverse experimental settings.

Main Results:

• The DDA algorithm achieves high accuracy in cross-subject and cross-session scenarios.
• Peak accuracy reached 91.08% and 92.89% on the SEED dataset.
• Peak accuracy reached 81.58% and 80.82% on the SEED-IV dataset.

Conclusions:

• Aligning EEG data within emotion categories is crucial for generalizable and discriminative features.
• The proposed DDA algorithm effectively addresses both global and local domain divergences.
• DDA significantly improves the performance of EEG emotion recognition systems in cross-domain settings.