Machine Learning-based Educational Analysis of Software Metrics for Software Defect Prediction

Abstract

Software defect prediction has become an important research area in software engineering education because early identification of defect-prone programs may support instructional intervention and improve programming quality. This study investigates the educational significance of software metrics for software defect prediction using machine learning and deep learning approaches. Rather than focusing solely on predictive performance, the study emphasizes educational interpretation by analyzing how software metrics reflect programming complexity, code quality, and defect-prone programming behavior. A dataset containing 10,885 software instances and 22 software metrics was analyzed using multiple feature groups, including McCabe metrics, Halstead metrics, operator-related metrics, and combined feature representations. Several machine learning models, including Logistic Regression, Extra Trees, and XGBoost, were evaluated using stratified cross-validation, while TabNet was employed as a deep learning model for tabular data analysis. Experimental results showed that the Extra Trees model using all software metrics achieved the best overall performance with an accuracy of 0.7904 and an F1-score of 0.4368. Complexity-related metrics such as cyclomatic complexity and branching density were identified as important indicators associated with debugging difficulty, reduced readability, and increased cognitive burden during programming tasks. The findings demonstrate that interpretable software metrics can provide meaningful educational insights for identifying defect-prone programming patterns and supporting educational monitoring. Overall, the proposed framework highlights that machine learning can function not only as a predictive tool but also as an analytical framework for understanding programming behavior and software quality in educational environments.