Composite and Multiple Kernel Learning for Brain Computer Interface

Abstract

High-performance feature engineering and classification algorithms are significantly important for motor imagery (MI) related brain-computer interface (BCI) applications. In this research, we offer a new composite kernel support vector machine (CKSVM) based method to extract significant common spatial pattern (CSP) feature components from multiple temporal-frequency segments in a data-driven manner. Furthermore, we firstly introduce a multiple kernel discriminant analysis (MKDA) method for MI EEG classification. The experimental results on BCI competition IV data set 2b clearly showed the effectiveness of our method outperforming other similar approaches in the literature.