A systematic review on AI based class imbalance handling in software defect prediction

• A Systematic Review on AI based Class Imbalance Handling in Software Defect Prediction. • It is a Systematic Literature Review (SLR) of 123 articles over 25 years during 2000 to 2025. • To conduct this review, the guidelines for SLR and PRISMA are followed. • Promise and NASA are highly imbalanced datasets and most popular ones. • Ensemble + Sampling techniques – most viable solution to CI. • AUC is the most used ones being insulated from the impact of imbalanced datasets. • Empirical comparisons are made to confirm statistically that CI is a major issue in SDP and must be handled carefully for accurate predictions. • A pathway for future researchers to make much informed decisions about the CIH in SDP using AI mechanisms backed by the proper evidence is led by this SLR. Software defect prediction is to predict the fault-prone modules assisting the development team to focus the testing efforts on the faulty modules, preserving the testing resources and leading to reduced costs of the product. A huge range of diversified AI based mechanisms are available to handle the class imbalance issue underlying the defect datasets that would hinder the performance of prediction models. This study assesses the timeline of evolution of AI applications in the domain of software defect prediction and culminates a Systematic Review over the period of 25 years from 2000 to 2025 dedicatedly focusing on class imbalance issue. To conduct this review, the PRISMA guidelines and statistical hypothesis testing approach are followed. The aim of this study is to assess the AI practices being applied to strengthen prediction models while alleviating the class imbalance. The study about the publication trend in the past 25 years reveals that majority of defect datasets (∼77%) suffer from class imbalances and data sampling is the most popular class imbalance handling technique. Machine learning techniques have been popular and now the paradigm is shifting towards deep models and transfer learning. Explainable AI and LLMs are now the names for defect prediction models over industry datasets. This review provides a holistic view about the defect datasets, learning models, evaluation criteria, class imbalance handling techniques that will pave guidelines and references to the future researchers and software engineers to achieve more reliable and effective software defect prediction models with appropriately adapting the AI techniques to handle the class imbalance conditions.