Development and Performance Assessment of Software Defect Prediction Using Machine Learning Approaches

Abstract

Abstract newlineSoftware defect prediction (SDP) is essential for improving software quality, reducing newlinecosts, and ensuring system reliability by detecting defect-prone modules early. Accurate newlinepredictions help prioritise testing, optimise resources, and minimise rework. However, newlineSDP faces challenges, such as high-dimensional metrics, noisy and redundant features, newlineand limited generalisation across datasets. To address these issues, this research develops newlineframeworks that combine dimensionality reduction, metaheuristic optimisation, and extreme newlinelearning machines. Techniques like PCA and PCA-LDA are applied to retain only the most newlinerelevant and discriminative features. Metaheuristic -optimisation algorithms are employed newlineto refine feature subsets and fine-tune model parameters for greater accuracy. ELM has been newlinechosen as the primary classifier because of its speed, scalability, and strong generalisation newlineabilities. The models are systematically evaluated on benchmark NASA datasets (CM1, newlineKC2, MC1, KC3, PC1, JM1). Comparative experiments show that optimisation-driven newlineELM models consistently outperform conventional classifiers, such as MLP, J48, AdaBoost, newlineRandom Forest, and Decision Tree. The aim of this thesis is to develop advanced machine newlinelearning frameworks that integrate dimensionality reduction, optimisation techniques, and newlineextreme learning machines to achieve robust and accurate software defect predictions. newlineThis research introduces four different frameworks that combine dimensionality reduction newlinemethods with advanced classifiers for accurate software defect prediction. newlineThe first contribution presents a model for predicting software defects based on extreme newlinelearning machines (ELM), enhanced through the use of dimensionality reduction techniques newlinesuch as PCA and PCA-LDA. These methods help eliminate redundant and irrelevant features newlinewhile preserving the most significant and discriminative attributes, thereby improving newlineprediction performance. When evaluated on NASA datasets, the PCA-LDA+ELM model newlineconsistently outperformed conventional classifiers, with the KC2 dataset yielding the best newlineresults using a reduced feature set. This established PCA-LDA+ELM as a strong baseline newlinefor further refinements in subsequent contributions. newlineThe second contribution introduces an ELM model integrated with the improved JAYA newline(IMJAYA) optimisation algorithm, combined with PCA and PCA-LDA for dimensionality newlinereduction. This approach refines feature subsets and optimises classifier parameters to newlinefurther enhance predictive accuracy. Evaluations on six NASA datasets demonstrated that newlinethe PCA-LDA+IMJAYA-ELM framework produced more effective results than standard newlinev newlineclassifiers, with the KC2 dataset again showing the strongest outcomes using fewer newlinefeatures. This highlighted benefits of incorporating optimisationtion techniques in the defect newlineprediction process. newlineThe third contribution proposes an intelligent defect prediction framework that newlineintegrates PCA-LDA with the improved grey wolf optimisation (IMGWO) algorithm newlineand ELM. By balancing exploration and exploitation in feature reduction, IMGWO newlineenhances the global search capability, ensuring that the most relevant features are newlinechosen. The resulting PCA-LDA+IMGWO-ELM model achieved superior performance newlineacross the NASA datasets, with the KC2 dataset delivering the most promising results. newlineThis contribution demonstrated the effectiveness of metaheuristic-driven optimisation for newlineimproving prediction accuracy and robustness. newlineThe fourth contribution presents an efficient software defect prediction framework newlinethat integrates PCA-LDA, improved particle swarm optimisation (IMPSO), and ELM. newlineBy effectively balancing exploration and exploitation during parameter tuning, IMPSO newlineenhances the optimisation process, enabling the classifier to perform reliably with a newlinereduced set of features. The proposed PCA-LDA+IMPSO-ELM model demonstrated newlinestrong predictive accuracy and robustness across NASA datasets, underscoring the value newlineof optimization-driven learning models for improving the effectiveness and efficiency of newlinedefect prediction systems. newlineIn summary, this thesis demonstrates that integrating dimension reduction with newlineadvanced optimisation-driven extreme learning machine (ELM) models significantly newlineimproves the accuracy, robustness, and efficiency of software defect prediction. Across newlinemultiple NASA datasets, the proposed frameworks consistently outperformed conventional newlineclassifiers, highlighting the importance of feature refinement and optimisation when newlinebuilding reliable prediction systems. The progression from PCA-LDA+ELM to newlineoptimisation-enhanced hybrids, such as IMJAYA-ELM, IMGWO-ELM, and IMPSO-ELM, newlineestablishes a comprehensive framework that reduces feature dimensionality and achieves newlinesuperior predictive performance. These findings give useful insights into the development newlineof intelligent, scalable, and resource-efficient models, offering practical implications for newlinereal-world software engineering and quality assurance practices. newlineKeywords; Software defect prediction; Dimensionality reduction; Extreme learning newlinemachine; Metaheuristic Optimisation; NASA Datasets. newline