Regularized Common Spatial Pattern with Incomplete Generic Learning for EEG Classification in Small Sample Setting

Abstract

Common spatial pattern (CSP) is based on sample covariance matrix estimation, whose estimation accuracy is severely overfit with small training sets. To address the drawback, the regularized CSP (R-CSP) was proposed that adds regularization information into the CSP learning process. In the algorithm, all samples of each generic subject were used for training sample covariance matrices. When only a part of the samples of each generic subject are allowed as generic training set, this R-CSP algorithm wouldn’t work. To solve this problem, an improved method is proposed in this paper. The new algorithm was applied to a brain-computer interface (BCI) data set containing five subjects and a mean improvement of 2.5% in classification rate was achieved.