EARLY STAGE BUG DETECTION AND TRIAGING USING MACHINE LEARNING

Abstract

Early stage bug prediction and triaging of software are essential to ensuring software reliability and minimizing downstream maintenance. With growing software systems, hand triaging does not work, is error-prone, and cannot be scaled. Recent progress in machine learning presents promising opportunities for automating these tasks by applying machine learning to learn historical software repository trends. This study explores various supervised machine learning methods for binary prediction of bug reports to facilitate early-defect prediction and triaging. Various classifiers such as ensemble methods, support-vector machines, and neural networks were created and tested on real-world bug databases. To deal with class imbalance, both the oversampling and undersampling methods were utilized and their effect on model performance was determined. The main performance metrics to be evaluated were accuracy, F1-score, and area under the receiver operating characteristic curve. Model comparison was to determine the most stable and consistent models in terms of these performance metrics, with special interest in how sampling strategies affected consistency of performance. The analysis explored model sensitivity to class imbalance and behavior pattern as the characteristics. It was found that ensemble techniques were very sensitive to sampling methods and outperformed regular classifiers when the issue of data imbalance was predominant. This study contributes to the literature a strong early bug prediction framework with machine learning that provides explicit model and sampling choice advice to improve performance, especially when dealing with imbalanced datasets. The new approach makes it simple to create smart automatic tools for bug triaging to aid software teams in maximizing defect management effectiveness and code quality at scale.