Software Defect Prediction Based on Machine Learning and Deep Learning Techniques: An Empirical Approach

Abstract

Software bug prediction is a software maintenance technique used to predict the occurrences of bugs in the early stages of the software development process. Early prediction of bugs can reduce the overall cost of software and increase its reliability. Machine learning approaches have recently offered several prediction methods to improve software quality. This paper empirically investigates eight well-known machine learning and deep learning algorithms for software bug prediction. We compare the created models using different evaluation metrics and a well-accepted dataset to make the study results more reliable. This study uses a large dataset collected from five publicly available bug datasets that includes about 60 software metrics. The source-code metrics of internal class quality, including cohesion, coupling, complexity, documentation inheritance, and size metrics, were used as features to predict buggy and non-buggy classes. Four performance metrics, namely accuracy, macro F1 score, weighted F1 score, and binary F1 score, are considered to quantitatively evaluate and compare the performance of the constructed bug prediction models. The results demonstrate that the deep learning model (LSTM) outperforms all other models across these metrics, achieving an accuracy of 0.87.