Graph Attention-Based Contextual Semantic Correlation Analysis for Software Fault Localization

Abstract

Although the advanced spectrum-based fault localization techniques introduce graph models to measure the faulty suspiciousness of entities, they still lack effective leverage of the contextual semantics of program units, making it difficult to reveal the complex causal relationships of bugs. Therefore, this paper proposes the Contextual Semantic Coupling Correlation Analysis for the Software Fault Localization (CSCFL), employing semantic embedding and coupling correlation of contextual associations based on contextual embedding and graph attention mechanism. Based on the test coverage graph representation, the pre-trained self-attention embedding model encodes the entity embeddings considering context associations, fully mining the contextual information among program entities to obtain contextual feature representation. CSCFL introduces the graph attention network to learn the context semantic coupling correlation between entities. By analyzing the causal relationships and semantic collaborations that lead to faults, the context semantics analysis module automatically weights and aggregates the influence of contextual entities to generate more accurate fault feature representations, thereby improving localization effectiveness. The experimental results show that, with respect to Top-1, Top-3, and Top-5 metrics, the proposed method outperforms the baseline methods by $7.59\%,4.73\%$, and $1.41\%$, respectively, and outperforms cutting-edge LLM-based methods by $9.10\%,13.71\%$, and $20.67\%$.