Comparing the classification performance of Bayesian linear discriminate analysis (BLDA) and support vector machine (SVM) in BCI P300-speller with familiar face paradigm

P300-speller which relies on P300-event related potential (ERP) is an important application of the BCI system. However, the accuracy and information transmission rate were relatively low for practical use. To solve the problem, researchers focused on two aspects of paradigms and classifiers. P300-speller with familiar face paradigm achieved a better performance. In addition, Bayesian linear discriminate analysis (BLDA) was successfully applied to classifier in the P300-speller with familiar face paradigm. However, BLDA needed the large scale of training set to ensure the accuracy of classification. Subjects gradually got sleepy during the period of training set collection, which resulted in an ambiguous unobvious difference between the attended and unattended characters, and affected the accuracy of classification. Support vector machine (SVM) is a strong classification method which achieves a very good classification in the case of small training set. However, whether SVM has a good performance on a familiar face paradigm or not remains unclear. In this study, we compared the classification performance of BLDA and SVM in BCI P300-speller with familiar face paradigm. In the condition in which the size of the training set is very small, P300-speller achieved a better performance using SVM classifier than using BLDA classifier. Along with gradually increasing of the size of the training set, the difference of classification performance between SVM and BLDA becomes small. Our results demonstrated the classification performance of SVM is better than that of BLDA in P300-speller with familiar face paradigm using a small training set.