Automatic recognition of electroencephalogram (EEG) signals plays a major role in epilepsy diagnosis and assessment. However, the recognition accuracy of conventional methods is usually not satisfactory because of the inconsistent distribution of training and testing data in practical applications. To overcome this problem, we used cross-domain mean joint approximation embedding (CMJAE) transductive transfer learning method to realize the knowledge transfer from the training data to the testing data by measuring the distribution difference between them. We combined the subspace learning and joint distribution to adapt the marginal and conditional distribution discrepancy. Our method was able to effectively learn a model for the testing data from training data with different distribution at a low computational complexity cost. On a public dataset, an ad-hoc cross-validation scheme of the proposed method exhibited that the average recognition accuracy, sensitivity, specificity of different states was 97.5%, 94.3%, 92.7% respectively, much better than conventional machine learning or deep learning methods, which may serve as a promising strategy for epileptic states recognition algorithms.