LocatorGraph: GNN Fault Localization for DL Programs
Novel GNN framework converting sequence model code into TraceGraphs to localize structural and runtime deep learning bugs (Empirical Software Engineering 2025).
LocatorGraph: GNN Fault Localization for Deep Learning Programs
Stack: Python, Graph Neural Networks, PyTorch Geometric, AST Analysis
Journal: Empirical Software Engineering (EMSE), 30(5):119, 2025
DOI: 10.1007/s10664-025-10666-6
Authors: Muhammad Anas Raza, Mohammad Wardat
Overview
Localizing bugs in deep learning code is notoriously difficult due to complex dataflow dependencies, dynamic tensor shapes, and silent runtime dimension mismatches.
LocatorGraph (LG) introduces a Graph Neural Network-based fault localization framework that parses sequence-based model code (RNNs, LSTMs, GRUs, Transformers, gMLP) into structured TraceGraphs to identify and pinpoint root causes of failure.
Empirical Evaluation (152 Sequence Models)
- Bug Identification: Achieved 89.46% AUC and 81.34% F1-score in classifying whether a sequence model implementation contains structural flaws.
- Feature Inspection: Achieved 91.19% accuracy and 89.68% F1-score on inspecting graph-based computation features.
- Faulty Node Pinpointing: Localized exact bug-causing nodes within computation graphs with 81.54% accuracy and 79.98% F1-score, consistently outperforming baseline spectra-based and static fault localization techniques.