Self-supervised Group Meiosis Contrastive Learning for EEG-Based Emotion Recognition

The progress in EEG-based emotion recognition has received widespread attention from the fields of human-machine interactions and cognitive science in recent years. However,the scarcity of high granularity and dense artificial labels contributes to the bottleneck of emotion recognition with high real-time and granularity. Neuroscience studies have discovered exploitable stimuli labels. It reveals that EEG samples recorded when subjects were triggered by the same stimuli share the same stimuli label. This paper proposed a Self-supervised Group Meiosis Contrastive Learning (SGMC) framework to exploit such stimuli labels for emotion recognition. The SGMC adopts a genetic inspired data augmentation method Meiosis. It achieves augmenting a group of samples sharing the same stimuli label to generate two augmented groups by pairing, cross exchanging, and separating. A group-projector-based model is adopted. The model achieves extracting a group-level representation by extracting individual representations from a group and aggregating them into a group-level representation. Contrastive learning is employed to maximize the similarity of group-level representations of augmented groups sharing the same stimuli label. The SGMC achieved the state-of-the-art results on the publicly available DEAP dataset with an accuracy of 94.72% and 95.68% in valence and arousal dimensions. Especially the SGMC shows more excellent performance on limited labeled sample learning. In addition, we verified the rationality of the framework design by control experiment and ablation study, and investigated the cause of the formation of good performance by feature visualization, and hyper parametric analysis. The code is provided publicly online