Adaptive semi-supervised graph contrastive representation learning for robust emotion recognition using electroencephalogram

Emotion recognition using electroencephalogram (EEG) signals remain challenging due to the brain’s complex and variable activity. Recent graph-based methods have significantly improved performance by modeling intricate EEG relationships. However, most existing approaches rely on static graphs that often fail to capture the dynamic, non-stationary nature of EEG and underutilize abundant unlabeled data, limiting robustness given EEG’s noisy nature and costly labeling process. To address these limitations, we propose ASGCRL, an adaptive semi-supervised graph contrastive representation learning framework. ASGCRL dynamically constructs EEG graphs based on evolving inter-sample relationships, mitigates abnormal connections through a novel adaptive augmentation strategy, and effectively leverages both limited labeled and abundant unlabeled data via consistency regularization. Extensive experiments on four public datasets demonstrate that ASGCRL outperforms state-of-the-art supervised models despite using less than 10% of labels. Additional analyses confirm its superior label efficiency and robustness in EEG-based emotion recognition.