Based on existing software defect data, it is possible to improve the efficiency of software testing and reduce the test cost by establishing the classification model to predict the software modules. Most machine learning based defect prediction researches are based on two-way decision method. Since software defect prediction can be seen as a kind of cost-sensitive learning problem, and the software data has continuous values, this paper proposes a classification method based on neighborhood three-way decision-theoretic rough set model. For ambiguous testing modules, compared with two-way decision methods, this method makes a deferment decision to reduce the misclassification rate. Experimental results on NASA software datasets show that the proposed method can get a higher classification accuracy and a lower misclassification cost.