A Motor Imagery Transfer Learning Approach Based on Riemannian Alignment and Distance Preservation to Local Mean

The classification models generally have poor generality in the Motor Imagery Brain-Computer Interface systems (MI-BCIs). Besides, the long training process limits the practicality of the system. Thus, shortening the long training process and improving the generalization of the system's model are very important for developing BCI systems. Transfer learning (TL) is currently one of the most commonly used strategies in BCI to reduce the calibration time and improve the adaptability of cross-subject learning systems, which can eliminate the dependence on model reconstruction. To address the lack of adaptability of MI-BCI systems in cross-subject learning, we propose a TL algorithm that combines Riemannian alignment with Distance Preservation to Local Mean (RA-DPLM), which utilizes the Riemannian distance affine invariance, effectively combines Riemannian geometry and TL to align the distributions of data on Riemannian manifolds between the target and source subjects to reduce the training time for cross-subject experiments. Meanwhile, the proposed method combined with the dimensionality reduction algorithm further reduces the calibration time for cross-subject experiments. Experimental results have demonstrated the effectiveness of the proposed RA-DPLM algorithm, showing that its integration with dimensionality reduction techniques significantly improves the cross-subject classification accuracy for classical left- and right- hand motor MI tasks.