Related Experiment Video
Updated: May 16, 2026
Conscious and Non-conscious Representations of Emotional Faces in Asperger's Syndrome
Published on: July 31, 2016
SFT-SGAT: A semi-supervised fine-tuning self-supervised graph attention network for emotion recognition and
Lina Qiu1, Liangquan Zhong2, Jianping Li2
1School of Artificial Intelligence, South China Normal University, Guangzhou, 510630, China; Research Station in Mathematics, South China Normal University, Guangzhou, 510630, China.
This study introduces a novel semi-supervised fine-tuning self-supervised graph attention network (SFT-SGAT) for improved cross-subject electroencephalogram (EEG) emotion recognition. The SFT-SGAT method achieves state-of-the-art accuracy and shows potential for assessing consciousness in patients with disorders of consciousness (DOCs).
Area of Science:
- Neuroscience
- Computer Science
- Biomedical Engineering
Background:
- Brain-computer interfaces (BCIs) rely on emotional recognition, but individual variability in electroencephalogram (EEG) signals and labeling challenges hinder cross-subject performance.
- Traditional methods struggle with the noise and complexity of EEG data, limiting their effectiveness in real-world applications.
Purpose of the Study:
- To develop a robust cross-subject EEG emotion recognition method that overcomes limitations of existing approaches.
- To enhance the generalization ability and accuracy of EEG-based emotion recognition models using limited labeled data.
Main Methods:
- Proposed a semi-supervised fine-tuning self-supervised graph attention network (SFT-SGAT) to model spatiotemporal EEG features.
- Employed self-supervised learning to mitigate signal noise and a semi-supervised approach for fine-tuning to improve generalization.
- Utilized graph structures to dynamically capture topological features within multi-channel EEG signals.
Main Results:
- Achieved state-of-the-art cross-subject emotion recognition accuracies of 92.04% on the SEED dataset and 82.76% on the SEED-IV dataset.
- Demonstrated high classification performance in healthy subjects (up to 95.84%) and successful application to patients with disorders of consciousness (DOCs), with some exceeding 60% accuracy.
- The SFT-SGAT model effectively combined supervised and unsupervised learning to maximize the utility of limited labeled EEG data.
Conclusions:
- The SFT-SGAT model significantly improves cross-subject EEG emotion recognition.
- The method shows promise for assessing emotional states and levels of consciousness in individuals, including patients with DOCs.
