Abstract: Machine-learning-based software defect prediction methods are received widely attention from the researchers in the field of software engineering.The defect distribution in the software can be analyzed by the defectprediction mo-del,so as to help the software quality assurance team to detect potential software errors and allocate test resources reasonably.However,most of the existing defect prediction methods are based on hand-crafted features such as line of code,dependency between modules and stack reference depth.These methods do not take into account the potential semantic features of the software source code and may result in poor predictions.To solve the above problems,this paper applied convolutional neural networks to mine the semantic features implicit in the source code.In the effective mining of source code semantic features,this paper used three-layer convolutional neural network to extract data abstract features.In terms of data imbalance processing,this paper adopted a cost-sensitive method,which gives different weights to positive and negative examples,and balances the impact of positive and negative examples on model training.In terms of experimental data sets,this paper selected multiple versions of the eight softwares in the PROMISE defect dataset,totaling 19 projects.In terms of model comparison,this paper compared the proposed cost-sensitive software defect prediction model based on convolutional neural network (CS-TCNN) with logistic regression and deep confidence network respectively.The evaluation metrics contain AUC and MCC,which are widely used in the field of defect prediction research.The experimental results demonstrate that CS-TCNN can effectively extract the semantic features in the program code,and improve the prediction effect of the software defect prediction model.

Key words: Convolutional neural network, Cost-sensitive, Semantic feature mining, Software defect prediction