Improving fault localization with semantic enhancement and decision fusion
Benyu Liu, Pan He · 2025
Automatic fault localization(FL) is a crucial step in software debugging tasks. In recent years, most research has incorporated the rich features of software into deep learning based fault localization methods, including both static and dynamic features. However, existing methods have yet to achieve satisfactory results. The main reason is that existing deep learning based methods use a single feature fusion approach to integrate different types of features, failing to fully utilize the information of the software system, which results in low fault localization accuracy. To address these issues, this paper proposes a fault localization method based on semantic enhancement and decision fusion(SEDeFL). SEDeFL employs a two-stage fusion to integrate different types of features in software. First, it uses codeBERT and GNN to extract the sequence-based and structure-based features of the code, integrating them to obtain enhanced semantics. Then, it employs a decision fusion method to integrate static features and runtime features of the software, maximizing the reduction of feature loss. Experiments on the benchmark dataset Defects4J demonstrate that our method can detect more faulty statements, achieving 54.1%, 32.4%, 10.1% and 6.82% higher Top-1 accuracy compared to the four best deep learning based methods, respectively. This indicates that our method exhibits good fault localization performance.