The data set originally has 64 EEG channels. We assumed that some of them were 
either redundant one another or irrelevant to the classification task. Therefore, 
we reduced the dimensions of this dataset directly in terms of original variables 
(channels) to improve the classification performance.



To identify those significant channels out of 64, we first applied our own 
feature subset selection (FSS) method, CLeVer, proposed in [1]. Then, each of 
EEG signals was transformed into a vector using the upper triangle of its 
correlation coefficient matrix, where only the selected channels were used 
and then fed into linear SVM classifier. That is, the vectorized EEG signals 
with the selected channels only were used for both the training and test dataset 
to train the linear SVM classifier and predict the labels of test data, 
respectively.



[1] Hyunjin Yoon, Kiyoung Yang, Cyrus Shahabi, Feature Subset Selection and 
Feature Ranking for Multivariate Time Series, IEEE Transactions on Knowledge 
and Data Engineering (TKDE) - Special Issue on Intelligent Data Preparation, 
2005, (To appear)