SOFTWARE DEFECT PREDICTION FOR PYTHON SOFTWARE USING ENSEMBLE LEARNING: EXPERIMENTAL ANALYSIS AND REVIEW OF SEQUENTIAL PREDICTION MODELS

Abstract

Software Defect Prediction is another sub field of Software Quality Assurance that is a core area and aims to identify which parts are more likely to be faulty during early stages of the development life cycle, allowing the teams to better allocate testing efforts on areas that are more likely to fail. Most of the previous work has been limited to C, C++, and Java programming languages using the standard NASA and PROMISE datasets, although some gaps exist. On the one hand, there is still little guidance for Python specific defect prediction, while on the other, there does not seem to be a clear guidance in the literature regarding which sequential deep learning architectures perform better. The two peer reviewed contributions contained within this thesis examine both of these concerns. This thesis aims to fill both of these gaps in two peer reviewed contributions. The first paper presents an empirical evaluation of five ensemble methods– XGBoost, LightGBM, CatBoost, Stacking Classifier, and Voting Classifier on a dataset derived from PyTraceBugs dataset of 638,568 class instances with 18 static code metrics and a binary defect label. Random Oversampling is employed in order to reduce the high degree of class imbalance (buggy instances are 0.038% of the total). The features extraction is done in two stages, resulting in a set of 15 features.The Stacking Classifier achieved the highest F1-score (0.6966) and MCC (0.6986), while the Voting Classifier attained the highest Precision (0.7632) and LightGBM achieved the best AUC-ROC (0.9455), indicating strong performance under extreme class imbalance. The second contribution is a systematic review of 36 peer-reviewed Seq-SDP studies (2018–2025), analyzing model architectures, datasets, evaluation methods, performance trends, and research challenges. The review highlights the dominance of hybrid and LSTM-based models, the widespread use of the PROMISE repository, and the strong focus on Java, emphasizing the need for Python-based research. The most efficient model is the CodeBERT-BiGRU model with an accuracy of 93%, F1 score of 88.9% and MCC of 0.85.