A Comparative Study of Classification Techniques and Imbalanced Data Treatment for Prediction of Software Faults

Abstract

Abstract Software Defect Prediction is one of the major challenges faced by software engineers across the world as software grows in size and function. It is the process of identifying error-prone modules in software before the testing phase, which helps with cost-cutting and saves time. The primary goal of this research is to compare the different data balancing techniques along with the popular classification models used for software fault prediction and optimize the best results. In this study, we have used the AEEEM dataset, along with mean value treatment and min-max scaling to pre-process data. Then dataset balancing is performed using class-weight-based, over-sampling, under-sampling, and hybridization techniques. The balanced datasets are now analyzed using 5 classification techniques: Random Forest Classifier, XGBoost, Support Vector Classifier, LightGBM, and Logistic Regression. Thus, a total of 25 combinations are accessed to find the best results using 10-fold cross-validation with f1-score and AUC as the performance metric. Further, the best methods are improved using feature selection. Finally, the best case is optimized using Optuna.