A FAULT DETECTION STRATEGY FOR SOFTWARE PROJECTS

Abstract

The existing software fault prediction models require metrics and fault data belonging to previous software versions or similar software projects. However, there are cases when previous fault data are not present, such as a software company's transition to a new project domain. In this kind of situations, supervised learning methods using fault labels cannot be applied, leading to the need for new techniques. We proposed a software fault prediction strategy using method-level metrics thresholds to predict the fault-proneness of unlabelled program modules. This technique was experimentally evaluated on NASA datasets, KC2 and JM1. Some existing approaches implement several clustering techniques to cluster modules, process followed by an evaluation phase. This evaluation is performed by a software quality expert, who analyses every representative of each cluster and then labels the modules as fault-prone or not fault-prone. Our approach does not require a human expert during the prediction process. It is a fault prediction strategy, which combines a method-level metrics thresholds as filtering mechanism and an OR operator as a composition mechanism.