Source file level software defect prediction framework

Abstract

Defect prediction techniques are used to address defective sections of source code in software products. Applying a defect prediction technique before proceeding to testing phase of software development helps the managers to allocate their resources more e ciently and most core e ectively such as time and e ort to test certain sections of the code. Defect predictors are useful tools to help project managers to plan test stage during the software development life cycle without compromising on the product quality. In this research we have taken software defect prediction as a two way classi cation problem. We have used machine learning techniques to construct our prediction model. One of the challenges in learning based models is the collection of data. In software engineering domain data collection is a major problem. Companies and researchers often struggle to nd out the right level of granularity in data collection: i.e. module / function level versus le / class level. In this research we have been motivated by the problem of right level of granularity. Our proposed models use the hierarchical structure information about the source code of the software product, in order to perform defect prediction for high level granularity such as source les (also called classes). We have run experiments on NASA, SoftLab and Eclipse datasets to validate our proposed model. Additionally we have also performed cost-bene t analysis to evaluate the net e ect of using our proposed model.