| Abstrakt: |
Software development projects are prone to bugs that can hinder the development process and delay software release. Among the most critical bugs are blocking bugs (BB), which impede further testing or development until they are resolved. Identifying BBs promptly can significantly reduce the time software products are held up due to these bugs. Researchers have developed predictive models using machine learning algorithms to identify BBs. However, seminal studies on blocking bug prediction did not perform a parameter optimization study or a comprehensive evaluation of their models, despite empirical evidence showing that hyperparameters (i.e., tuning) directly influence the behavior of machine learning techniques and significantly impact the performance of the prediction model being trained. To provide a manual configuration for achieving optimal predictive performance and avoiding using default parameter values, this paper investigates the effect of parameter tuning on blocking bug prediction models by replicating a seminal research. We analyze six different open-source projects, including Chromium, Eclipse, NetBeans, Gentoo, OpenOffice, and Free Desktop, using twelve machine-learning algorithms operating under various parameter settings. The results demonstrate that our models can predict BBs with an AUC value of up to 88.5% and precision values of 81.8%, outperforming state-of-the-art approaches. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink