 In this paper, we proposed a novel semi-supervised cross-session EG Emotion Recognition Framework called JCSFE. The framework consists of two main components, one, exploring label common and label specific EG features, two, exploiting the data local invariance property. We conducted experiments on the seed4 and seedv data sets and compared our results with other state-of-the-art methods. Our results show that JCSFE outperforms all other methods in terms of accuracy and robustness. Furthermore, we provide a quantitative analysis to identify the label common and label specific EG features. This article was authored by Yongpeng, Honggang Lu, Junwali, and others.