Scholarlink

<p>Software defect prediction (SDP) is essential for maintaining software quality, especially in critical systems like telecommunications. However, the lack of transparency in deep learning models hampers their practical application. This paper introduces an explainable deep learning framework for software defect prediction (SDP), combining a residual-shuffle network (RSN) with two model-agnostic interpretability tools—Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME)—to enhance the explainability of predictions using explainable artificial intelligence (XAI). The RSN architecture is designed to capture complex software patterns through residual learning and channel shuffling, while SHAP and LIME provide interpretable insights into individual predictions and global feature importance. We evaluate the framework on public benchmark datasets (PROMISE/NASA) and industrial datasets from the telecommunications sector. Results show that the RSN + XAI model outperforms traditional models in terms of accuracy, precision, recall, and F1-score. Additionally, user feedback from industrial practitioners reveals that the interpretability of the model aids in resource allocation, debugging, and meeting regulatory compliance. This work demonstrates the importance of model transparency in SDP and paves the way for future research on real-time feedback mechanisms and the integration of natural language features to further improve model explainability in software engineering contexts.</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

Explainable deep learning for software defect prediction: a residual-shuffle network approach with SHAP and LIME

  • 期刊论文

  • 发表日期：

    2026年1月12日

查看全文
  • Samira Keramat Talatapeh,
  • Mohammad Reza Ebrahimi Dishabi,
  • Hamed Pezeshki,
  • Halimeh Madadi

摘要

Software defect prediction (SDP) is essential for maintaining software quality, especially in critical systems like telecommunications. However, the lack of transparency in deep learning models hampers their practical application. This paper introduces an explainable deep learning framework for software defect prediction (SDP), combining a residual-shuffle network (RSN) with two model-agnostic interpretability tools—Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME)—to enhance the explainability of predictions using explainable artificial intelligence (XAI). The RSN architecture is designed to capture complex software patterns through residual learning and channel shuffling, while SHAP and LIME provide interpretable insights into individual predictions and global feature importance. We evaluate the framework on public benchmark datasets (PROMISE/NASA) and industrial datasets from the telecommunications sector. Results show that the RSN + XAI model outperforms traditional models in terms of accuracy, precision, recall, and F1-score. Additionally, user feedback from industrial practitioners reveals that the interpretability of the model aids in resource allocation, debugging, and meeting regulatory compliance. This work demonstrates the importance of model transparency in SDP and paves the way for future research on real-time feedback mechanisms and the integration of natural language features to further improve model explainability in software engineering contexts.