Machine Learning Approaches for Software Fault Diagnosis: Evaluating Decision Tree and KNN Models

This research study explores the predictive modeling of software errors by employing machine learning techniques, using the PC1 dataset extracted from PROMISE Software engineering repository. In this work, it is proposed to investigate the attributes related with dependability and quality of software that can be used for an accurate prediction model in detecting significant software problems, aiming at improvements in fault-prone detection. We evaluated two machine learning models: Decision Tree and K-Nearest Neighbors (KNN) using several performance indicators. The results indicate that KNN achieves higher AUC and accuracy, but Decision Tree model offers desirable balance between precision-recall trade-offs which is applicable for multiple condition-specific fault prediction scenarios. In hindsight, this work highlights the importance of selecting machine learning methodologies and evaluation metrics that are tailored to software failure prediction use-cases.