Kai-Min Chung Email: b88061@csie.ntu.edu.tw Tzu-Kuo Huang Email: b89034@csie.ntu.edu.tw Chih-Jen Lin Email: cjlin@csie.ntu.edu.tw Computer Science and Information Engineering National Taiwan University, Taipei, Taiwan Description: Since we are developers of SVM software, we use SVM as the classifer for this problem. Observing that the dataset has a few training trials with a huge number of features, feature selection becomes the main challenge of this dataset. After trying several methods, we use the following simple strategy: 1. Train the data by linear SVM to get a model (w, b). 2. Select several hundreds of features with the largest w_i. 3. Train the new data with nonlinear SVM. Surprisingly, after doing model selection on SVM parameters, this method gets nearly 100% cross-validation accuracy. Although we think that overfitting may occur, we decided to submit the result using this strategy.