Relation Between the Best Subject's Set of Tasks and the Classifier With the Best Performances on This Set

This study continues another BCI research in which we explored the positive significant impact that could be obtained in tasks classification accuracy only by selecting the individual most appropriate set of mental tasks. In this paper, the same EEG database was used; namely, for each investigated subjects we classified both a classical reference set of four mental tasks (count, letter, math and rotate) and the best set of four mental tasks, determined individually, in the former research. Twenty different classification systems were tested, each of them belonging to one of the following families: decision trees, support vector machines, nearest neighbor classifiers, and ensemble classifiers. In this study two goals were intended: (a) to investigate, in what degree, a distinct relationship could be established between each subject's specific (given) set of tasks and the classifier, and (b) to explore if and how the subjects' individual-selected sets of tasks could impact on the performance ranking of the classifiers. The study results indicated that no particular relationship appeared, but two other important outcomes were emphasized: (a) the consistency of the considerable improvement generated in the classification accuracy by changing, appropriately and for each individual apart, only the set of the analyzed mental tasks, and (2) the enhancing effect of linearization achieved in the input space and only for the subject's proper set of tasks.