Related Experiment Videos
Prototypical graph based deep label propagation with semantic augmentation for cross-subject and cross-session EEG
Yufang Dan1,2,3,4, Li Zhu5, Di Zhou6
1School of Business Intelligence, Zhejiang Institute of Economics and Trade, Hangzhou, China.
Frontiers in Neuroscience
|July 22, 2026
Summary
This study introduces Prototypical Graph-based Deep Label Propagation with Semantic Augmentation (PGDLP) to improve electroencephalogram (EEG) emotion recognition across different users and sessions. PGDLP enhances generalization by semantically augmenting data and using graph-based label propagation for more accurate emotion classification.
Area of Science:
- Neuroscience
- Machine Learning
- Affective Computing
Background:
- Electroencephalogram (EEG)-based emotion recognition struggles with generalization due to signal non-stationarity and individual differences.
- Existing methods often fail to adapt effectively across different subjects or sessions, limiting practical applications of brain-computer interfaces.
Purpose of the Study:
- To propose a novel unsupervised domain adaptation framework, Prototypical Graph-based Deep Label Propagation with Semantic Augmentation (PGDLP), to address generalization challenges in EEG emotion recognition.
- To enhance the universality and accuracy of emotion recognition systems by bridging the domain gap between different EEG datasets.
Main Methods:
- PGDLP integrates prototypical semantic augmentation, prototype-graph deep label propagation, and prototypical alignment.
- Semantic augmentation uses source-domain statistics to create class-wise multivariate normal distributions for target-domain features.
- An adaptive similarity graph refines pseudo-label quality, combined with linear projection and exponential moving average (EMA) for dynamic updates.
Main Results:
- PGDLP achieved superior recognition accuracy and generalization performance on benchmark datasets (SEED, SEED-IV, DEAP) across various cross-domain protocols.
- The framework demonstrated strong robustness against label noise, stable hyperparameter performance, and fast convergence.
- Outperformed state-of-the-art methods in unsupervised cross-domain EEG emotion recognition.
Conclusions:
- PGDLP offers a promising solution for unsupervised cross-domain EEG emotion recognition, significantly improving generalization.
- The study provides valuable insights into domain adaptation techniques for physiological signal analysis.
- The proposed framework enhances the reliability and applicability of affective brain-computer interfaces.
