<p>Domain adaptation is a crucial factor in EEG emotion recognition as it allows the transfer of knowledge from labeled source domains to unlabeled target domains with different data distributions. However, traditional domain adaptation methods require labeled source domain data, which is often unavailable due to privacy concerns in EEG emotion recognition. Consequently, source-free unsupervised domain adaptation has emerged as a challenging problem. To address this issue, this study proposes a contrastive self-supervised learning (CSSL) framework for online calibration of EEG recognition without using source data. The CSSL framework comprises two steps. In the first step, shared and personalized generator and classifier models are trained using contrastive learning techniques to make shared and personalized features more suitable for clustering and generating pseudo-labels. In the second step, pseudo-labels generated by Gaussian mixture model are used to supervise the training of the target model. Meanwhile, to make the pseudo-labels more reliable, mutual information maximization is applied on enhanced adjacent samples. The experimental results indicate that the proposed method outperforms state-of-the-art methods, achieving an accuracy of 89.2% and 61.6% on the SEED and DEAP datasets, respectively.</p>