Kernel Nonnegative Matrix Factorization with Constraint Increasing the Discriminability of Two Classes for the EEG Feature Extraction

Nonnegative matrix factorization (NMF) is an algorithm for blind source separation. It has been reported that the use of kernel NMF (KNMF) is a particularly feasible way to extract the features of a motor-imagery related EEG spectrum, which is often used in brain-computer interfaces (BCI). A BCI system enables users to control electrical devices without their hands or feet, and often requests to tell user's intention from motor-imagery related EEG features. In other words, a classification of the EEG signals reflecting the user's intentions is required. In this research, a constraint is placed on the KNMF to increase the discriminability between two classes, widening the difference between their spectral EEG energies. To evaluate the proposed method, the IDIAP database, which contains the motor-imagery related EEG spectrum of three subjects, was adopted for the discrimination between two classes. As a result, the classification accuracy when using the proposed constraint was approximately 78% on average, which is 4% higher than that obtained by KNMF without a constraint.