Pascal Challenge: Linear Support Vector Machines

We participate in the linear SVM Track of the Pascal Large Scale Learning Challenge at ICML 2008. We consider the LIBLINEAR package, which can handle L1- and L2-loss linear SVMs. The L1-SVM solver implemented in LIBLINEAR employes a coordinate descent method to solve the dual problem. This method is very useful for large sparse data with a huge number of instances and features. However, most data sets of this challenges have a quite small number of features. To work on the competition data, we slightly modify LIBLINEAR.