Algorithms Description:
1. Preprocessing Techniques:
   1.1 Down-sample the data to 125 Hz;
   1.2 Use linear regression to remove EOG artifacts before the subsequent data processing;
   1.3 Filter the re-referenced signals in an 8-25Hz band with a causal linear IIR band-pass filter;
   1.4 Use a algorithm which based on a recursive channel elimination to select the channels.
2. Designing a ensemble muti-class classifer from 3 two-hierarchy classifier:
   2.1 divide the trainset into 3 subsets, each subset containing 24 x 4 trials; 
   2.2 Doing Common Spatial Pattern (CSP) for feature extration;
   2.3 Design the first hierarchy two-against-two classifer by Support Vertor Machine (SVM);
   2.4 Design two second hierarchy SVMs as one-against-one classifers.
3. Combine the classification outputs from 3 two-hierarchy classifiers by a voting strategy.
4. Down-sample the continuous classification outputs from 125Hz to 5 Hz. Consequently, there are 37 outputs in a    single trial.

Result Data Format:
Results for the nine subjects are saved in nine Matlab-format files, prelabel_A0X.
Each file contains the continuous classification outputs for the whole dataset. Its size is 288*37 for all subjects, where 288 is the number of trials for a subject and 37 is the number of time points within a trial.

Names and Affiliations:
SONG Wei [1]   EMAIL TO: songw57@gmail.com
WU Jin [2]   EMAIL TO: wujin03@gmail.com
ZHANG Jiacai [1]   EMAIL TO: jiacai.zhang@bnu.edu.cn
[1]. College of Information Science and Technology, Beijing Normal Unviersity
[2]. National Key Laboratory for Cognitive Neuroscience and Learning, Beijing Normal Unviersity