Data augmentation effects using borderline-SMOTE on classification of a P300-based BCI
Taejun Lee, Minju Kim, Sung-Phil Kim
Abstract
In this study, we addressed a problem of imbalance in the size of event-related potentials (ERPs) between target and nontarget stimulation events, which is intrinsic to the odd-ball paradigm used in P300-based brain-computer interfaces (BCIs). Specifically, we investigated whether data augmentation could remedy this problem and improve BCI performance. We investigated a data augmentation technique, borderline-Synthetic Minority Over-sampling Technique (SMOTE). We focused on the effects of data augmentation on users with poor BCI performance. The EEG data were obtained from experiments with the P300-based BCI system developed for controlling 3 home appliances (Lamp, Door lock, Bluetooth speaker), where the classifier was designed by a support vector machine (SVM) and a convolutional neural network (CNN). As a result, although Borderline-SMOTE did not significantly change the overall BCI performance, it significantly improved the performance of poor performers. This suggests that data augmentation can offer an effective way to increase the performance of users illiterate to P300-based BCIs.