Feature Clustering and Ensemble Learning Based Approach for Software Defect Prediction

Abstract

Objective:: Defects in delivered software products not only have financial implications but also blemish the reputation of the organisation and lead to wastage of time and human resource. This paper aims to detect defects in software modules. Methods:: Our approach sequentially combines SMOTE algorithm to deal with class imbalance problem, K - means clustering algorithm to obtain a set of key features based on inter-class and intra-class coefficient of correlation and ensemble modelling to predict defects in software modules. After cautious examination, an ensemble framework of XGBoost, Decision Tree and Random Forest is used for prediction of software defects owing to numerous merits of ensembling approach. Results:: We have used five open-source datasets from NASA Promise Repository for Software Engineering. The result obtained from our approach has been compared with that of individual algorithms used in ensemble. A confidence interval for the accuracy of our approach with respect to performance evaluation metrics namely Accuracy, Precision, Recall, F1 score and AUC score has also been constructed at a significance level of 0.01. Conclusion:: Results have been depicted pictographically.