Tagging multimedia content in terms of emotions can improve the efficiency of user recommendations. Using a Brain-Computer Interface of emotion based on Electroencephalogram (EEG) is a reliable method to annotate the emotions evoked by content. However, the subject-independent models’ generalization and re-usability often degrade due to individual differences in EEG data. Hence, in this paper, we proposed a novel unsupervised domain adaptation method incorporating a variant of MMD (Maximum Mean Discrepancy) called Categorical MMD (CMMD) with a curriculum pseudo-labeling module to tackle the above mentioned issue. Unlike the original MMD, CMMD considers the relationship between data belonging to the same category from the source and target domain so that CMMD can align the distribution more than fine-grained. Besides, to mitigate the influence of false pseudo labels on CMMD and final classification, we proposed a curriculum pseudo labeling module that leverages reliable pseudo-target samples and labeled source samples to train the classifier. The experiment was conducted on the SEED dataset to verify our method. The experimental results showed that the proposed method achieves significant performance.