Abstract: Cost-sensitive learning is popular during the process of classification. Most researches focus on two costs for building cost-sensitive decision trees, such as, misclassification costs, test costs. In this study, a novel splitting attributes criterion is proposed firstly. And a test strategy combining discount costs for decreasing the misclassification cost is presented with missing values in test set after the cost-sensitive decision tree are constructed with missing values in training sets. Finally, the experimental results show our method outperform the existed methods in terms of the decrease of misclassification cost.