Scholarlink

<p>Identifying change-prone areas in software is vital for effective maintenance, as it reduces effort and costs while ensuring high-quality software. Achieving real-time predictions of software change-proneness requires training machine learning (ML) models. However, training ML models on high-dimensional open-source datasets is computationally intensive. Therefore, selecting an optimal set of representative features prior to training is essential to reduce computational demands and improve model performance. To this end, existing studies have applied various feature reduction methods; however, the resulting reduced feature sets can sometimes be less suitable or challenging for software practitioners to interpret effectively. To address this, we propose an end-to-end framework utilizing an explainable AI (XAI) method– SHapley Additive explanations (SHAP)– to identify a minimal, meaningful set of predictors. We tested this approach on six open-source software datasets and statistically compared its feature selection capabilities against three traditional methods using six ML algorithms. Our results indicate that machine learning models built using features selected by SHAP achieved higher AUC scores and demonstrated more consistent performance across most of the evaluated datasets. The average AUC obtained through 10-fold cross-validation using SHAP-selected features was <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="500_2025_10751_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="80" /> </InlineMediaObject> <EquationSource Format="TEX">\(0.61 \pm 0.05\)</EquationSource> </InlineEquation> at the 95% confidence level. In addition to improved performance, SHAP-based models offered enhanced interpretability, enabling software practitioners to devise more effective maintenance strategies while monitoring a reduced set of features.</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-AI-assisted feature selection for software change-proneness prediction

  • 期刊论文

  • 发表日期：

    2025年6月1日

查看全文
  • Ankit Rajpal,
  • Megha Khanna,
  • Naman Singhal

摘要

Identifying change-prone areas in software is vital for effective maintenance, as it reduces effort and costs while ensuring high-quality software. Achieving real-time predictions of software change-proneness requires training machine learning (ML) models. However, training ML models on high-dimensional open-source datasets is computationally intensive. Therefore, selecting an optimal set of representative features prior to training is essential to reduce computational demands and improve model performance. To this end, existing studies have applied various feature reduction methods; however, the resulting reduced feature sets can sometimes be less suitable or challenging for software practitioners to interpret effectively. To address this, we propose an end-to-end framework utilizing an explainable AI (XAI) method– SHapley Additive explanations (SHAP)– to identify a minimal, meaningful set of predictors. We tested this approach on six open-source software datasets and statistically compared its feature selection capabilities against three traditional methods using six ML algorithms. Our results indicate that machine learning models built using features selected by SHAP achieved higher AUC scores and demonstrated more consistent performance across most of the evaluated datasets. The average AUC obtained through 10-fold cross-validation using SHAP-selected features was \(0.61 \pm 0.05\) at the 95% confidence level. In addition to improved performance, SHAP-based models offered enhanced interpretability, enabling software practitioners to devise more effective maintenance strategies while monitoring a reduced set of features.