Working set selection is an important step in SMO for training support vector machine (SVM). Faced with C-SVM, Fan Rong-En proposed a method, which used second-order approximate information to select working set, and indicated that it had higher rate than the maximal violating pair. Based on this method, faced with weighted support vector machine (W-SVM) this paper proposes a training algorithm, which uses second-order approximate information to select working set. At the same time, two data preprocessing methods are proposed for existing weight knowledge and non-existing weight knowledge. Experiments indicate that the methods not only ensure precision, but also improve training rate highly