Dual regularized spatial-temporal features adaptation for multi-source selected cross-subject motor imagery EEG classification

Feature adaptation plays crucial roles in the calibration process of motor imagery brain computer interfaces (MIBCIs). Due to the temporal varying and spatial coupling characteristics in MI-electroencephalograph (EEG), recently proposed cross-subject MI-EEG classification methods have suffered from patterns collapse and erroneous labels accumulation, evenly lower efficiency. To address these fundamental limitations, this paper proposes a novel method to represent Spatial-Temporal Features Adaptation based on Dual Regularizations (DRSTFA), and perform a simplest multi-source domain samples selection during adaptation. Specifically, the covariance centroid alignment is applied as the preprocessing, and then common spatial-temporal pattern (CSTP) is represented for aligned MI-EEG samples. Finally, the dual regularizations of cross-domain graph preservation and target domain discriminability strengthen have been incorporated into joint distribution adaptation framework for CSTP feature adaptation. The proposed method has been systematically benchmarked on three BCI competition MI-EEG datasets, and its classification performance surpasses several state-of-the-art methods. Moreover, it effectively captures meaningful temporal varying and spatially coupled features with parameter insensitivity. Our method therefore provides a novel calibration choice for newly subjects participating in MI-BCIs.