A Transformer Framework Based on Self-Supervised Contrastive Learning for EEG-based Emotion Recognition

Deep learning has achieved a great success in various fields, such as image classification and so on. But its excellent performance tends to rely on a large amount of labeled data that are hard to collect and manual annotation, especially for EEG signals. Self-supervised learning (SSL), as a solution, can discover structure in unlabeled data and learn global representations. Therefore, for emotion recognition, we propose a Transformer Framework based on Self-supervised Contrastive Learning (TFSCL). We design positive and negative pairs based on temporal and spatial characteristic of EEG to pre-train a general model for crossing subjects. We conduct a convolution-based patch embedding framework and a self-attention Transformer encoder to learn local and global representations of EEG signals. We adopt experiments on our proposed TFSCL using the DEAP and SEED emotion datasets, and perform better accuracy in cross-subject tasks. Experimental results suggest that self-supervised learning may play a greater role in deep learning models of EEG signals.