Algorithms Description: 1. Prepeocessing. 1.1 Down-sample the data to 125 Hz; 1.2 Select the electrodes (channels) around the C3 and C4 for each subject; 1.3 Select the time sections among the motor imagery course for each subject; 1.4 Filter the signals in an 8-25 Hz band IIR filter£» 1.5 Doing EOG removal with the auto-regression method for the trials which contains artifacts. 2. Designing a two-hierarchy muti-class classifer by boosting ensemble learning method. 2.1 Doing Common Spatial Pattern (CSP) for feature extration; 2.2 Design the first hierarchy Support Vertor Machine (SVM) as two-against-two classifer (class 1 and 2 against class 3 and 4); 2.3 Design two second hierarchy SVMs as one-against-one classifers. 3. All parameters were tested with 2-fold cross-validation. Result Data Format: The result is a continuous classification outputs for each sample in the form of class labels for each subject. Its size is N*T, where N is the number of the trials and T is the total number of the sample points in one trial. Here, N = 288 and T = 938 as we have down-sampled the data to 125 Hz. Names and Affiliations: WU Jin [1] EMAIL TO: wujin03@gmailcom CHEN Guangming [2] EMAIL TO: ctycheer@yahoo.com.cn SONG Wei [2] EMAIL TO: songw57@gmail.com ZHANG Jiacai [2] EMAIL TO: jiacai.zhang@bnu.edu.cn [1]. National Key Laboratory for Cognitive Neuroscience and Learning, Beijing Normal Unviersity [2]. College of Information Science and Technology, Beijing Normal Unviersity