Linear classification is an old topic, but recently its large-scale training has become a popular research issue. One reason is that training and testing linear classifiers is much more efficient than nonlinear classifiers with kernels. In this talk, we explain that linear classification is useful for some problems such as document classification. We then discuss optimization methods for fast training and testing. Most of these methods are included in our ongoing development of the software LIBLINEAR -- a library for for large linear classification. In LIBLINEAR, we have solvers for logistic regression and support vector machines under different regularizations. For example, solvers for L1-regularized problems produce sparse representations useful for many new applications. We finally discuss extensions of LIBLINEAR for some special large-scale scenarios.