Data set IVa Group Information: Department of Biomedical Engineering, Southeast University, Nanjing, 210096, P. R. China Supervisor: Wang Feng, Ph.D. (wangfeng.list@seu.edu.cn) Team Member: Liu jianmin, Gong Kainan, Wang Xiaonan, Tang Liyan Correspondance: yrmfwxn@yahoo.com.cn or cnyzgkn@163.com Algorithm description The algorithm includes three steps: channel selection, feature extraction, and classification. Channel Selection First, we eliminated useless channels based on energy, and then selected channels with correlation coefficients. About 26 channel is selected. Feature Extraction In the session of feature extraction, we used two methods including Common Spatial Patterns (CSP) and Common Spatial Subspace Decomposition (CSSD) and obtained two sets of features. 1. Spatial Feature Extraction(use CSP) First we filtered the signals with three FIR filters of 0-7Hz, 9-13Hz and 18-33Hz respectively and obtained three sets of data. Then, we applied CSP to these three data sets and get spatial features. 2. Time Feature Extraction(use CSSD) After the filtering with a FIR filter 0f 0-7Hz, we used CSSD and obtained the time features. Classification In this session, we used the classifier based on SVM algorithms, both linear and nonlinear. For each subject, we used the former half data as training sets while the latter half as testing set in order to evaluate these algorithms.We obtained the best algorithms and their parameters that have the highest correct classifying rate for each subject. We re-trained all the selected algorithms by using all the five data files and get the results.