Development of Predictive Software Defect Detection Models Using Machine Learning Techniques across Different Phases of the Software Development Life Cycle

Abstract

Software defects remain one of the most critical challenges in software engineering, significantly affecting software reliability, maintainability, performance, and overall development cost. Traditional defect detection approaches primarily rely on manual inspections and testing activities performed during the later stages of the Software Development Life Cycle (SDLC), often leading to delayed fault identification and increased correction expenses. To address this limitation, this research proposes the development of predictive software defect detection models using machine learning techniques across different phases of the SDLC, enabling early and accurate identification of defect-prone software components. Recent literature consistently supports the value of early defect prediction using machine learning, particularly across requirement, design, coding, testing, and maintenance phases.  The proposed framework integrates multiple supervised machine learning algorithms, including Decision Tree, Random Forest, Support Vector Machine, Logistic Regression, and Artificial Neural Networks, to analyze historical software metrics and phase-specific defect indicators. Input features include requirement volatility, design complexity metrics, coupling and cohesion measures, cyclomatic complexity, lines of code, defect density, and change frequency. The collected data are preprocessed through normalization, missing-value handling, and feature selection techniques to improve model performance and reduce dimensionality. The predictive models are trained and evaluated using benchmark software defect datasets such as NASA PROMISE, Eclipse, and open-source repository datasets, ensuring model generalizability and robustness. Performance evaluation is conducted using widely accepted metrics including accuracy, precision, recall, F1-score, ROC-AUC, and Matthews Correlation Coefficient (MCC). It is expected that ensemble learning methods, particularly Random Forest and hybrid models, will outperform conventional classifiers in terms of predictive accuracy and early defect detection capability. Current systematic reviews show these algorithms remain among the strongest performers in software defect prediction studies. This study contributes a comprehensive intelligent defect prediction framework that supports proactive quality assurance, risk reduction, and optimized resource allocation throughout the SDLC. The proposed approach is expected to minimize maintenance costs, improve software quality, and facilitate data-driven decision-making in modern software development environments.