 Transfer learning, TL, has been successfully applied to seizure detection problems, where it is able to leverage knowledge from other subjects to improve performance. This paper proposes a novel approach called cluster embedding joint probability discrepancy transfer, CJT. Firstly, the joint probability distribution discrepancy is minimized to reduce the distribution shift between the source and target domains. Secondly, a clustering algorithm is employed to identify the best subjects for transferring knowledge. Finally, a correlation alignment-based source selection metric, SSC, is introduced to select the most suitable subjects for transfer. Experimental results show that CJT outperforms existing methods and demonstrates its effectiveness in improving seizure detection accuracy. This article was authored by Xiao Nanchui, Zhou Enchao, Jiao Pinglai, and others.