Industrial applications of artificial intelligence in software defects prediction: Systematic review, challenges, and future works

Abstract

Software defect prediction is a constant challenge in industrial software engineering and represents a significant problem for quality and cost in software development worldwide. The purpose of this study is to gain a deeper understanding of the quartiles, countries, keywords, techniques, metrics, tools, platforms or languages, variables, data sources, and datasets used in software defect prediction. A comprehensive search of 45 articles from 2019 to 2023, using 5 databases (Scopus, ProQuest, ScienceDirect, EBSCOhost, and Web of Science), was conducted following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis) methodology. Results show that 60.00 % of the studies were carried out in 2023, and 68.89 % of journals were in the Q1 and Q2 quartiles. The most common techniques were Support Vector Machine (42.22 %) and Random Forest (35.56 %). The most commonly used evaluation metrics were Accuracy and F1-Score (68.89 %). Python was the main programming language (35.56 %), with Kilo (thousands) of lines of code (31.11 %) and Cyclomatic complexity (26.67 %) as key variables. Finally, NASA's Metrics Data Program Data Repository was the most used data source (31.11 %) with a dataset ranging from a minimum of 759 instances and 37 attributes to a maximum of 3579 instances and 38 attributes from 5 projects: CM1, MW1, PC1, PC3, and PC4. This systematic review provides scientific evidence on how machine learning algorithms aid in predicting software defects and improving development processes. In addition, it offers a detailed discussion by identifying trends, limitations, successful approaches, and areas for improvement, providing valuable recommendations for future research.