Research On Software Defect Prediction Algorithm Based On Cost-sensitive Learning

Software defect prediction(SDP)can accurately identify whether there is a defective software programs in a software module by designing a robust machine learning model,and then provide guidance for the rational allocation of the test resources and the improvement of the software reliability.The cost of misclassifying a defective software program is much higher than that of misclassifying a non-defective,which is the cost-sensitive learning problem.Also,the amount of instances in target project is rare,and we hope to improve the performance of the learning model on the target project by utilizing instances from other software projects whose data distribution is different,which is the cross-project problem in SDP.Regarding the issue above,the contributions of this thesis in two aspects can be summarized as follows:(1)To solve the cost-sensitive learning problem in SDP,this thesis proposes a cost-sensitive dictionary learning(CS-DL)method for software defect prediction.By jointly optimizing the cost-sensitive dictionary and coefficients in a unified cost-sensitive dictionary learning framework,the discriminative dictionary will be achieved for feature encoding of test data.In the experimental part,we have conducted extensive experiments on twenty-five software projects from four benchmark datasets of NASA,AEEEM,Re Link and Jureczko.The results,in comparison with ten state-of-the-art software defect prediction methods,demonstrate the effectiveness of proposed CS-DL method for software defect prediction.(2)For cross-project software defect prediction problem,we propose a cost-aware based graph convolutional neural network(CAGCN).First,the code of source project and target project is parsed and converted into numeric vectors that can express grammatical structure and semantic information.Then,a cost-aware layer is designed.Finally,the cost-aware layer is embedded in the graph convolutional neural network in a layer-wise way to learn the cost-sensitive graph neural network parameters.The extensive experiments on ten software projects with different defect ratios.In comparison with seven representative software defect prediction methods,the experimental results demonstrate the effectiveness of the proposed CAGCN method.