Assessment of machine learning algorithms for determining defective classes in an object-oriented software

Abstract

Software defect prediction is a well renowned field of software engineering. Determination of defective classes early in the lifecycle of a software product helps software practitioners in effective allocation of resources. More resources are allocated to probable defective classes so that defects can be removed in the initial phases of the software product. Such a practice would lead to a good quality software product. Although, hundreds of defect prediction models have been developed and validated by researchers, there is still a need to develop and evaluate more models to draw generalized conclusions. Literature studies have found Machine Learning (ML) algorithms to be effective classifiers in this domain. Thus, this study evaluates four ML algorithms on data collected from seven open source software projects for developing software defect prediction models. The results indicate superior performance of the Multilayer Perceptron algorithm over all the other investigated algorithms. The results of the study are also statistically evaluated to establish their effectiveness.