 This paper proposes a new method called optimal graph coupled semi-supervised learning, OGSSL, for EEG Emotion Recognition. The proposed method combines adaptive graph learning and Emotion Recognition into a single objective, improving the label indicator matrix of unlabeled samples to directly obtain their emotional states. Additionally, the key EEG frequency bands and brain regions in Emotion Expression are automatically recognized by the projection matrix of OGSSL. Experimental results on the C4 data sets show that OGSSL achieves excellent average accuracies of 76.51%, 77.08% and 81.29% in three cross-session Emotion Recognition tasks, demonstrating its effectiveness and discriminative EEG feature selection and identifying the gamma frequency band, the left slash right temporal, prefrontal, and central parietal lobes as being more correlated with the occurrence of emotions. This article was authored by Yong Peng, Feng Zhenjin, Wanzin Kong, and others. We are article.tv, links in the description below.