Semi-supervised EEG emotion recognition with Discriminative Graph Transformer Model

Abstract

The complexity and vulnerability to the noise of the electroencephalogram (EEG) result in uneven data distribution, leading to performance degradation in emotion recognition. To address this issue, we propose the Discriminative Graph Transformer Model (DGTM). The DGTM is designed to capture the low-dimensional manifold structures inherent in high-dimensional EEG features. This approach enables the DGTM to distinguish data points from different classes with enhanced accuracy and informativeness. To demonstrate the superior performance of our model to the limit, we conduct experiments in a semi-supervised setting, utilizing only a small amount of labeled data ranging from 2.40% to 3.99% of the total dataset. The results showed that the pro-posed DGTM outperforms state-of-the-art models in the semi-supervised setting despite the lack of available ground truth data.