A Study of Ensemble Models for Defect Prediction from Class Diagram

Abstract

Software defect prediction in the early stages of the Software Development Life Cycle (SDLC) is crucial to reducing project cost and ensuring the implementation’s success. Existing methods for software defect detection in a project rely on the implementation or testing phases of the SDLC, based on the source code. While relatively few studies have focused on identifying defects in the design phase of the SDLC, these approaches primarily employ machine learning or deep learning methods to detect and classify suspect code segments or classes in static diagrams as defective or clean. This study utilizes 24 model-based metrics extracted via SDMetrics, including structural and objectoriented design features derived from UML class diagrams. To enhance classification performance, this study introduces an ensemble machine learning model with different techniques (stacking, voting) that combine multiple machine learning models. Specifically, we compare ensemble models with different ensemble techniques to the individual models in terms of accuracy, precision, recall, F-measure, and AUC by utilizing a large dataset called the Unified Bug Dataset, comprising five publicly available sub-datasets. Experimental results show that the ensemble model with the stacking ensemble method outperformed other ensemble models and the individual classifiers (RF, XGBoost, ET) in terms of AUC.