Software Defect Prediction Using Fuzzy Support Vector Regression

With development of information technology, computers are popular for their large storage capacity and fast computing speed. Computers play an important role in aerospace, transportation, finance, entertainment, health care and education. With increasing dependence on computers, possibility that software defects result in crisis increases. Software defects have a major impact on software quality. Software defect prediction techniques can predict the distribution and number of defects, which can help improve software quality and provide insurance against software defect.This paper studies fuzzy support vector machine based software defect prediction distribution. The main work and contributions are as follows:1. By analyzing the number of defects in modules of MIS and RSDIMU software metrics datasets, a fuzzy membership function fit for module defect number prediction is proposed. Based on this fuzzy membership function, Fuzzy Support Vector Regression is applied to predict number of defects in each software module. Comparison with Support Vector Regression is carried out. Experiments show that FSVR gets better results in the subset of fault-prone modules, whose Mean Squared Error of both software metrics datasets MIS and RSDIMU are lower than SVR. However SVR gets better MSE in the subset of fault-free modules.2. To get lower MSE when using SVR and FSVR, two commonly used software metrics dataset pre-process methods: principal component analysis and genetic algorithm are compared. Experiments show that genetic algorithm can select appropriate metrics subset for different software metrics datasets and gets better mean squared error than using principal component analysis in both methods support vector regression and fuzzy support vector regression.3. To get lower MSE for FSVR on condition that small impact is made on MSE of fault-prone modules, we adopt support vector classification to pre-process test sets to divide them into fault-prone modules and fault-free modules, and then FSVR and SVR are used for predicting number of defects in fault-prone and fault-free modules respectively. Experiments show that this method has small impact on MSE of fault-prone modules and increases the number of defects module prediction accuracy compared to using FSVR to predict the overall test set.