Score Normalization of Ensemble SVMs for Brain-Computer Interface P300 speller

Abstract

Brain-computer interface (BCI) P300 speller can be used as a powerful aid for severely disabled people in their everyday life. The character recognition using P300 speller involves two stages for classification. First stage is to detect the P300 signal and second one is to determine the right character from the detected P300. Classification of P300 is a challenging task in character recognition process. Ensemble of classifiers is a robust method for classification as it reduces the classifier variability. In multiclassifier system the averaged score can be effected by one classifier as the score of different classifiers are not in the same level. To reduce the effect of one classifier, the score of the each classifiers are normalized. The proposed method includes different score normalization techniques for ensemble of SVMs (ESVM) for classification. Here min-max normalization, Z-score normalization and median and median absolute deviation (MAD) normalization techniques are used. The proposed algorithms have been evaluated on data set II of the BCI Competition III. It is observed that the performance of the proposed normalization technique is better compared to the earlier reported techniques for 5th and 15th epoch to classify different characters