Towards a software defect proneness model: feature selection

Abstract

This article is focused on improving static models of software reliability based on using machine learning methods to select the software code metrics that most strongly affect its reliability.The study used a merged dataset from the PROMISE Software Engineering repository, which contained data on testing software modules of fiveprograms and twenty-onecode metrics. For the prepared sampling, the most important features that affect the quality of software code have been selected using the following methods of feature selection: Boruta, Stepwiseselection, Exhaustive Feature Selection, Random Forest Importance, LightGBM Importance, Genetic Algorithms, Principal Component Analysis, Xverse python.Basing on the voting on the results of the work of the methods of feature selection, a static (deterministic) model of software reliability has been built, which establishes the relationship between the probability of a defect in the software module and the metrics of its code. It has been shown that this model includes such code metrics as branch countof a program, McCabe’s lines of codeand cyclomatic complexity, Halstead’s total number of operators and operands, intelligence, volume, andeffort value.A comparison of the effectiveness of different methods of feature selection has been put into practice, in particular, a study of the effectof the method of feature selection on the accuracy of classification using the following classifiers: Random Forest, Support Vector Machine, k-Nearest Neighbors, Decision Tree classifier, AdaBoost classifier, Gradient Boosting for classification. It has been shown that the use of any method of feature selection increases the accuracy of classification by at least ten percentcompared to the original dataset, which confirms the importance of this procedure for predicting software defects based on metric datasets that contain a significant number of highly correlated software code metrics.It has been found that the best accuracy of the forecast for most classifiers was reachedusing a set of features obtained from the proposed static model of software reliability. In addition, it has been shown that it is also possible to use separate methods, such as Autoencoder,Exhaustive Feature Selection and Principal Component Analysis withan insignificant loss of classification and prediction accuracy.