COMPARATIVE ANALYSIS OF CLASSIFICATION AND ENSEMBLE METHODS FOR PREDICTING SOFTWARE FAULT PRONENESS USING PROCESS METRICS

Abstract

Various researchers have worked in the subject of software defect prediction to group the modules into defective or non-defective classes. But most of the previous studies done in this field utilize static code metrics to find the predicted value. The principal motive of this study is to evaluate the impact of process metrics on fault prediction performance using various classification techniques and ensemble techniques. In this study, we have analyzed the prediction performance of several classification and ensemble techniques based on three models: models that solely contain process metrics, models that solely contain static code metrics, and models containing different combinations of both metrics. In other terms, we can say these three models work as independent variables and dependent variables are actual bug values. We have used Naive Bayes classifiers, Logistic Regression, Support Vector Machines, K-Nearest Neighbors, and Decision Trees for implementation, and data sets are collected from publicly available repositories. We have also used four ensemble techniques: Stacking, Voting, Bagging, and Boosting to evaluate the impact of process metrics on fault prediction performance. We have also analyzed which process metrics give the best result among all selected process metrics. We have analyzed the prediction performance based on AUC (Area under ROC) performance measure and we have also used Friedman test with Nemenyi post hoc test to check whether the predictive performance of various classification techniques and ensemble techniques differ significantly. The result of this study shows that the use of process metrics in fault prediction gives effective results. In most of the cases, NR metric is effective when combined with static code metrics. If we consider combined model of 2 process metrics with static code metrics then combined model of NR, NDC metric with static code metrics gives effective result. If we consider combined model of 3 process metrics with static code metrics then combined model of NR, NDC, NDPV metric with static code metrics gives effective result.