BCI competition IV, data set I



 For this dataset, we used two steps of common spatial pattern. And solve the problem as 4-class classification problem: Left/Right(Foot)/Inter-trial/Inter-session. On the training set, data from 1s~4s after the cue was labeled as the cued motor imagery movements, and 5s~8s was labeled as the inter-session. 0s~1s and 4s~5s was ignored as transition state. And remaining regions was labeled as inter-session.


1st step) Classifying inter-session region.
 Mixing up inter-session and inter-trial as 'no cue' region deteriorates separation of inter-trial region between cues. So we firstly classified test set into [inter-session] and [cue/inter-trial] region. Short inter-session fragments was ignored becuase inter-session should have massive time length.
(Actually, this process properly worked only for data set F.)

2nd step) Classifying [Left/Right(foot)/Inter-trial]
 We solved this problem as 3-class classification by the CSP method, too. From Left and Right(foot) data sets, common spatial patterns were extracted. (without regarding inter-trial data). And few spatial filters which is corresponding to the largest and smallest eigenvalues was selected. And additionally, filters of corresponding eigenvalue is close to 0.5 were selected to extracting commonly observed pattern from both of conditions. By using these filters, the covariance matrix of each data set was projected into new feature space and trained by LIBSVM with a linear kernel function.

Artificial data :
I think Data set E (and probably G) is artificial data set. The result of CSP showed too perfact sptial patterns which denoting C3 and C4. And its corresponding eigenvalues were too dominant than eigenvalues from other noisy patterns.

reference)
Chih-Chung Chang and Chih-Jen Lin, LIBSVM : a library for support vector machines, 2001. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm