In emotion recognition research, EEG, as a direct response to brain activity, can objectively reflect a person's emotional state. However, the non-smoothness and other characteristics of EEG signals make it more difficult to collect a large number of labelled EEG samples, thus limiting the effectiveness and generalisation performance of EEG emotion recognition methods to a certain extent. To address the above problems, a semi-supervised low-rank representation of the EEG emotion recognition method is proposed. Firstly, a regression form objective function is designed using the estimated labels of a small number of labelled EEG samples as a way to effectively estimate the labels of unlabelled samples. Second, a -drag-and-drop technique is used to ensure label-to-label separability, and in addition, low-rank constraints are imposed on the slack labels to improve their intra-class tightness and similarity. Thirdly, a class neighbourhood graph is incorporated into the proposed method as a way to capture the local neighbourhood information of all EEG sample data. The proposed method is examined on the SEED-IV and SEED-V datasets, and the results show that the proposed method has better performance on the EEG emotion recognition problems.