Evaluation of Supervised Machine Learning for Software Defect Detection

Abstract

Software defect detection can support testing teams by identifying modules that are more likely to contain faults before full inspection or regression testing begins. This article presents a comparative evaluation of supervised machine learning algorithms for software defect detection using static, process, and historical software metrics. The study compares Logistic Regression, Decision Tree, Naive Bayes, Support Vector Machine, Random Forest, k-Nearest Neighbors, and Gradient Boosting under a common preprocessing, training, validation, and evaluation framework. Instead of relying only on overall accuracy, the evaluation focuses on defective modules detected, missed defective modules, false alarms, defect recall, review burden, and model stability across data splits. The results show that Gradient Boosting provides the strongest defect recall and lowest missed-defect count, while Random Forest offers the best balance between high detection performance and controlled false alarm burden. The proposed comparison supports practical classifier selection based on testing priorities, inspection capacity, and defect escape risk.