Semi-supervised Regularized Discriminant Analysis for EEG-Based BCI System

Brain-Computer interface (BCI) is a new technique which allows direction connection between human and computer or other external device. It employs the classification of event-related potential to control the equipment rather than using language or limb movement. Currently, one of the most important issues for ERP-based BCIs is that the ERP classification performance degrades when the training number of samples is small. In order to solve this problem, semi-supervised regularized discriminant analysis (SRDA) was proposed to extract features and classify ERP patterns by integrating semi-supervised learning and regularization approach. The labeled data was used to maximize the separability between different classes and calculate the within-class covariance matrix by regularization. The labeled and unlabeled data were employed to construct the penalty term by neighbor graph. Our proposed approach was evaluated on the BCI Competition Challenge Dataset and the simulation results indicated that it achieved a better accuracy than the traditional algorithms.