Automated Software Bug Severity Classification: A Comparative Performance Analysis of ML and DL Models

Abstract

In order to improve software quality, optimize resources, and minimize system failure, it is crucial to properly classify the severity of software issues. A comprehensive plan for automated bug severity classification using a weighted combination of source code metrics, class imbalance handling using the SMOTE algorithm, and a comparison of ML and DL models is proposed in this paper. A number of models were evaluated using the usual performance metrics, such as ROC-AUC, F1-score, recall, accuracy, and SVM, CNN, and LSTM. The experimental results demonstrate that SMOTE significantly improves recall and F1-score by increasing the detection of minority classes, particularly important faults. The capacity to reflect more complex trends in software measurements was proved by LSTM, which outperformed all other models with an accuracy of 89.2% and F1-score of 0.86. Research also provides further data by means of stability, robustness, and class-wise analyses, which demonstrate that DL models are better than conventional ML models. The suggested solution might make the problem triage process easier, better, and more automatic in real-world development settings. This might lead to faster debugging, better choices, and more dependable goods.