A comparison between different Multiclass Common Spatial Pattern approaches for identification of motor imagery tasks
Cecilia Lindig-León, Laurent Bougrain · HAL (Le Centre pour la Communication Scientifique Directe) · 2014
Common SpaMal PaOerns (CSP) is a feature extracMon method suited for two-class problems.However, there are some alternaMves to apply it for mulMclass tasks by using a group of ensemble classifiers that divide the problem into different binary classificaMon tasks, from which the final decision is inferred as the combinaMon of their responses.Nevertheless, there is another approach to extend CSP for mulMple classes to a one-step-method by approximaMng the joint diagonalizaMon of their covariance matrices [1] (fig.1).In this study, in order to idenMfy whether CSP by Joint Approximate DiagonalizaMon (JAD) represents an outperforming alternaMve to the standard mulMclass CSP alternaMves, four different methods were applied on dataset 2a used in BCI compeMMon IV [2] (fig.2).