A Comprehensive Evaluation of Machine Learning Techniques for Predicting Software Bugs

Software Bug Prediction (SBP) is essential to control overall software quality. As systems become complex, faults increase, and early prediction is needed to enhance quality and reduce expenses. This paper provides a technical study of thirteen Machine Learning (ML) techniques for predicting software bugs by utilizing NASA's Promise Repository datasets. The study examines the effectiveness of these classification techniques focusing on evaluation metrics such as Precision, Recall, F-measure, Accuracy, Error Rate, FPR, FNR, ROC, and MCC. The results indicate that Random Forest (RF) achieved the highest accuracy on PC1 (94.96%) and CM1 (96.60%), while Multilayer Perceptron (MLP) scored 85.6% on KC1, and Logistic Regression (LR) got 86.25% on KC2. Regarding error rates, RF had the lowest on PC1 (0.050), LR on KC2 (0.137). For FPR, SVM, MLP, LR, RBF, ANN, and AdaBoost excelled on PC1 (0.00). For FNR, Decision Tree (DT) performed best on PC1 (0.611), KC2 (0.518), and LDA on CM1 (0.500).