Defect Prediction Framework Using Neural Networks for Software Enhancement Projects

Abstract

So far, various approaches have been proposed for effective and accurate prediction of software defects, yet most of these approaches have limited adoption in practice. The objective of this paper is to provide a framework which is expected to be more user-friendly, effective and acceptable for predicting the defects in multiple phases across software enhancement projects. This communication describes a process of applying computational intelligence technologies, in particular neural networks in formulating defect prediction models early in the software development life cycle. A series of empirical experiments are carried out based on input and output measures extracted from 50 'real world' project subsystems. In order to increase the adoption and make the prediction framework easily accessible to project managers, a graphical user interface (GUI) based tool has been designed and implemented that allows input data to be fed easily. The proposed framework uses historical data for training model and as a result provides a defect range (minimum, maximum) based output instead of a definite defect count based output. This is done in view of the fact that exact-count prediction has less probability of being correct as compared to range based predictions. The defect predictions can be used for taking informed decisions including prioritizing software testing efforts, planning additional round of code reviews, allocating human and computer resources, planning for risk mitigation strategy and other corrective actions. The claim of effectiveness of proposed framework is established through results of a comparative study, involving the proposed framework and some well-known models for software defect prediction.