 In this study, we proposed a novel self-supervised learning method to improve the efficiency of resources usage for EEG-based emotion classification. The proposed method first transforms the raw EEG data into different representations and then uses a multitask CNN to recognize these representations simultaneously. Afterwards, the convolutional layer network is frozen and the fully connected layer is reconstructed as an emotion recognition network. Finally, the EEG data with effective labels are used to train the emotion recognition network. Experiments conducted from the data-scaling perspective using the CEDE and DEAP data sets demonstrated that the proposed self-supervised learning method outperforms the conventional fully supervised model. This article was authored by Xinyu Wang, Yulian Ma, Jared Kamen, and others.