Improved Label Space Alignment for Motor Imagery Classification

Motor imagery is one mode of brain-computer interface technology that can reflect person's motor intention. However, the EEG signals vary significantly over time and individuals, making it difficult to get the ideal sample data. In order to expand the sample data and make better use of multidomain data, this paper proposes a method called Improved Label Space Alignment (ILA), combined with heterogeneous transfer learning when the label space is heterogeneous. The core idea is to align with the class center of the source domain and the target domain, which includes three key steps: clustering, class matching and data alignment. We used three classification methods in binary classification heterogeneous scenario to verify our method. The results show that the improvement of ILA for the three classification methods was 20.88 %, 10.69% and 5.59%, and the effect is more obvious when the target domain labeled data is less.