Fig. 3

Performances on training data. Testing on training data (data from each CV fold that was also used to train on). Top: Stekelenburg&Vroomen data. Middle: BCI data, subjet A. Bottom: BCI data, subjet B. The methods are grouped by type such that first four methods plotted are the unsupervised decomposition methods, followed by the four heuristic supervised decomposition methods. The next four methods are the supervised manifold methods, which are followed by the two methods performing decomposition and classification in one step. Finally, the six methods that produce fewer features for classification are plotted again with 9 and 25 components