Scholarlink

<p>Deep learning models, particularly sequence-based models (SBMs) like Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs), Transformers, and patch-based architectures, are crucial for intelligent software systems. However, like any software, SBM applications are prone to bugs. Bug patterns in SBMs are distinguished from traditional techniques primarily due to their inherent black box, data-driven nature, and sequential dependencies. Moreover, currently used methods, although working for generic deep neural network (DNN) structures, require specialized expertise and are not directly applicable to the unique structures and requirements of SBMs. To address these challenges, we propose <i>LocatorGraph</i> (LG), a novel graph neural network-based framework designed to identify the root causes of faults in SBMs. To detect and localize faults in SBMs, we convert SBM code to TraceGraphs, which are then analyzed by <i>LocatorGraph</i>. The framework utilizes a graph neural network(GNN)–based architecture to not only identify bugs but also to precisely locate faulty features within the TraceGraphs. By leveraging this information, <i>LocatorGraph</i> can also pinpoint faulty nodes in SBMs through an evaluation of 152 diverse models, including generated buggy models, <i>LocatorGraph</i> outperforms existing methods for localizing faults, showing robustness to identify potential problems with an AUC of 89.46% and F1 score of 81.34%. On Graph-Based Feature Inspection, LG gives an accuracy of 91.19% and an F1-score of 89.68%. In the final phase, LG outperforms all baselines in locating bug causing nodes in the graph with an accuracy of 81.54% and an F1-score of 79.98%. We also provide details of the improvement observed in the models after fixing bugs using the <i>LocatorGraph</i>-based approach.</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

Graph neural network for fault localization in sequence-based models

  • 期刊论文

  • 发表日期：

    2025年6月2日

查看全文
  • Muhammad Anas Raza,
  • Mohammad Wardat

摘要

Deep learning models, particularly sequence-based models (SBMs) like Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Gated Recurrent Units (GRUs), Transformers, and patch-based architectures, are crucial for intelligent software systems. However, like any software, SBM applications are prone to bugs. Bug patterns in SBMs are distinguished from traditional techniques primarily due to their inherent black box, data-driven nature, and sequential dependencies. Moreover, currently used methods, although working for generic deep neural network (DNN) structures, require specialized expertise and are not directly applicable to the unique structures and requirements of SBMs. To address these challenges, we propose LocatorGraph (LG), a novel graph neural network-based framework designed to identify the root causes of faults in SBMs. To detect and localize faults in SBMs, we convert SBM code to TraceGraphs, which are then analyzed by LocatorGraph. The framework utilizes a graph neural network(GNN)–based architecture to not only identify bugs but also to precisely locate faulty features within the TraceGraphs. By leveraging this information, LocatorGraph can also pinpoint faulty nodes in SBMs through an evaluation of 152 diverse models, including generated buggy models, LocatorGraph outperforms existing methods for localizing faults, showing robustness to identify potential problems with an AUC of 89.46% and F1 score of 81.34%. On Graph-Based Feature Inspection, LG gives an accuracy of 91.19% and an F1-score of 89.68%. In the final phase, LG outperforms all baselines in locating bug causing nodes in the graph with an accuracy of 81.54% and an F1-score of 79.98%. We also provide details of the improvement observed in the models after fixing bugs using the LocatorGraph-based approach.