A survey on software fault detection based on different prediction approaches

Abstract

One of the software engineering interests is quality assurance activities such as testing, verification and validation, fault tolerance and fault prediction. When any company does not have sufficient budget and time for testing the entire application, a project manager can use some fault prediction algorithms to identify the parts of the system that are more defect prone. There are so many prediction approaches in the field of software engineering such as test effort, security and cost prediction. Since most of them do not have a stable model, software fault prediction has been studied in this paper based on different machine learning techniques such as decision trees, decision tables, random forest, neural network, Naïve Bayes and distinctive classifiers of artificial immune systems (AISs) such as artificial immune recognition system, CLONALG and Immunos. We use four public NASA datasets to perform our experiment. These datasets are different in size and number of defective data. Distinct parameters such as method-level metrics and two feature selection approaches which are principal component analysis and correlation based feature selection are used to evaluate the finest performance among the others. According to this study, random forest provides the best prediction performance for large data sets and Naïve Bayes is a trustable algorithm for small data sets even when one of the feature selection techniques is applied. Immunos99 performs well among AIS classifiers when feature selection technique is applied, and AIRSParallel performs better without any feature selection techniques. The performance evaluation has been done based on three different metrics such as area under receiver operating characteristic curve, probability of detection and probability of false alarm. These three evaluation metrics could give the reliable prediction criteria together.