Research and Appalication of Software Defect Predictionn based on BP-Migration learning

Abstract

Software Defect Prediction has been an important part of Software engineering research since the 1970s. This technique is used to calculate and analyze the measurement and defect information of the historical software module to complete the defect prediction of the new software module. Currently, most software defect prediction model is established on the basis of the same software project data set. The training date sets used to construct the model and the test data sets used to validate the model are from the same software projects. But in practice, for those has less historical data of a software project or new projects, the defect of traditional prediction method shows lower forecast performance. For the traditional method, when the historical data is insufficient, the software defect prediction model cannot be fully studied. It is difficult to achieve high prediction accuracy. In the process of cross-project prediction, the problem that we will faced is data distribution differences. For the above problems, this paper presents a software defect prediction model based on migration learning and traditional software defect prediction model. This model uses the existing project data sets to predict software defects across projects. The main work of this article includes: 1) Data preprocessing. This section includes data feature correlation analysis, noise reduction and so on, which effectively avoids the interference of over-fitting problem and noise data on prediction results. 2) Migrate learning. This section analyzes two different but related project data sets and reduces the impact of data distribution differences. 3) Artificial neural networks. According to class imbalance problems of the data set, using artificial neural network and dynamic selection training samples reduce the influence of prediction results because of the positive and negative samples data. The data set of the Relink project and AEEEM is studied to evaluate the performance of the f-measure and the ROC curve and AUC calculation. Experiments show that the model has high predictive performance.