Abstract: Support Vector Machine (SVM) as a machine learning method has rigorous theoretical basis and been used widely in engineering practices. How to improve the convergence speed of SVM is the hot research topic. Sequential Minimal Optimization (SMO) algorithm is an important decomposition method for SVM, in particularly, the working set selection algorithm is the key part for SMO. In most cases, the working set selection is selected directly based on the KKT conditions, but in some situations, the conventional selection method can't avoid the shortcoming resulting from random selection, even can't satisfy with the KKT conditions. In this paper, the condition of working set selection is improved by rewriting the KKT conditions, and a set of new steps for minimizing SMO sub-problem is proposed accordingly. The improved method is used to analyze the users' online behaviors, by comparing with the existing methods. It is validated that the new working set selection algorithms can reduce the learning time and accelerate the convergence speed of support vector machine, and improves the computational accuracy.