HGFE: A plug-and-play heterogeneous graph feature enhancement framework for software fault localization
TL;DR
The Heterogeneous Graph Feature Enhancer (HGFE) is proposed, an interface-preserving feature enhancement framework for downstream fault-localization models that consume feature matrices or feature vectors that consume feature matrices or feature vectors.
Abstract
Software fault localization is essential for reducing debugging effort and improving software quality, yet existing graph-based methods often focus on downstream model design while leaving weak input features and mixed relation semantics insufficiently addressed. This paper proposes the Heterogeneous Graph Feature Enhancer (HGFE), an interface-preserving feature enhancement framework for downstream fault-localization models that consume feature matrices or feature vectors. HGFE constructs sample-level multi-relational heterogeneous graphs over methods, statements, tests, regression tests, and mutants, and applies relation-specific message passing to transform spectrum- and mutation-based weak features into structure-aware representations. The enhanced features preserve the original downstream interface, allowing existing ranking models to use them without architectural modification when their input format is feature-based. Experiments on the Defects4J benchmark, covering Lang, Cli, and Math, show consistent improvements in Mean Fault Rank, Mean Reciprocal Rank, Normalized Discounted Cumulative Gain, and Top-k accuracy. Additional analyses indicate that explicit relation modeling, bidirectional propagation, residual connections, and sample-level batch isolation contribute to the observed gains. These results suggest that improving feature quality before downstream ranking is an effective and portable way to strengthen graph-based fault localization systems.
