Scholarlink

<p>Software fault prediction is a crucial task that aims to identify defective and erroneous software modules. Identifying faults in the software ensures smooth implementation and also eases the maintenance of the software in the future. This study aims to analyse Bayesian methods, which have emerged as a robust approach for the prediction of software faults, as these methods are capable of handling complex and high-dimensional data. This work has investigated the advanced Bayesian approaches for software fault prediction, focusing on Hierarchical Bayesian Neural Networks (HBNN), Bayesian Gaussian Process Regression with ARD kernel, Bayesian Deep Ensembles with posterior aggregation, and a Bayesian Mixture Model for fault-prone clustering. The performance of these Bayesian models is compared against a strong non-Bayesian baseline, Support Vector Regression (SVR).The study uses 42 versions of 11 open-source object-oriented software systems, covering both intra-release and cross-release prediction scenarios. Model performance is evaluated using Area Under the ROC Curve (AUC), Matthews Correlation Coefficient (MCC), and F1-score, providing a balanced assessment of discrimination capability, predictive correlation, and classification effectiveness. Furthermore, Box-plot analysis, the Wilcoxon signed-rank sum test, the Mann–Whitney U test, the K-S Anderson test, and the Rank-sums test have also been implemented for the statistical assessment of the given Bayesian methods to benchmark the effectiveness of the Bayesian methods. The study highlights the fact while SVR achieves strong discriminative performance, HBNN consistently offers the best balance between robustness, predictive reliability, and uncertainty-aware modeling, particularly in cross-release scenarios. The statistical analysis confirms that the observed improvements of hierarchical Bayesian models are significant, establishing them as a reliable and effective solution for software fault prediction under both stable and evolving software environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A statistical approach for software faults prediction using Bayesian methods

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Banoth Veeranna,
  • Sripriya Roy Chowdhuri,
  • Ragini Mishra,
  • Manjari Gupta

摘要

Software fault prediction is a crucial task that aims to identify defective and erroneous software modules. Identifying faults in the software ensures smooth implementation and also eases the maintenance of the software in the future. This study aims to analyse Bayesian methods, which have emerged as a robust approach for the prediction of software faults, as these methods are capable of handling complex and high-dimensional data. This work has investigated the advanced Bayesian approaches for software fault prediction, focusing on Hierarchical Bayesian Neural Networks (HBNN), Bayesian Gaussian Process Regression with ARD kernel, Bayesian Deep Ensembles with posterior aggregation, and a Bayesian Mixture Model for fault-prone clustering. The performance of these Bayesian models is compared against a strong non-Bayesian baseline, Support Vector Regression (SVR).The study uses 42 versions of 11 open-source object-oriented software systems, covering both intra-release and cross-release prediction scenarios. Model performance is evaluated using Area Under the ROC Curve (AUC), Matthews Correlation Coefficient (MCC), and F1-score, providing a balanced assessment of discrimination capability, predictive correlation, and classification effectiveness. Furthermore, Box-plot analysis, the Wilcoxon signed-rank sum test, the Mann–Whitney U test, the K-S Anderson test, and the Rank-sums test have also been implemented for the statistical assessment of the given Bayesian methods to benchmark the effectiveness of the Bayesian methods. The study highlights the fact while SVR achieves strong discriminative performance, HBNN consistently offers the best balance between robustness, predictive reliability, and uncertainty-aware modeling, particularly in cross-release scenarios. The statistical analysis confirms that the observed improvements of hierarchical Bayesian models are significant, establishing them as a reliable and effective solution for software fault prediction under both stable and evolving software environments.