Software Defect Prediction with Bayesian Approaches

Abstract

Software defect prediction is an important area in software engineering because it helps developers identify and fix problems before they become costly and hard-to-fix bugs. Early detection of software defects helps save time and money in the software development process and ensures the quality of the final product. This research aims to evaluate three algorithms to build Bayesian Networks to classify whether a project is prone to defects. The choice is based on the fact that the most used approach in the literature is Naive Bayes, but no works use Bayesian Networks. Thus, K2, Hill Climbing, and TAN are used to construct Bayesian Networks. On the other hand, three public PROMISE data sets are used based on McCabe and Halstead complexity metrics. The results are compared with the most used approaches in the literature, such as Decision Tree and Random Forest. The results from different performance metrics applied to a cross-validation process show that the classification results are comparable to Decision Tree and Random Forest, with the advantage that Bayesian algorithms show less variability, which helps engineering software to have greater robustness in their predictions since the selection of training and test data do not give variable results, unlike Decision Tree and Random Forest.