Automated Fault Localization Framework Using Program Dependency Analysis

Abstract

Automated fault localization reduces debugging effort by ranking program elements that are most likely to contain defects. This study proposes an Automated Dependency-Guided Fault Localization framework, called ADGFL, which combines execution spectra, control dependencies, data dependencies, method calls, variable-state divergence, and exception propagation. The framework first builds a hybrid execution dependency graph, removes weakly related nodes through failure-oriented slicing, and then ranks suspicious methods and statements using failure coverage contrast, dependency-path centrality, state divergence, control impact, and propagation distance. ADGFL was evaluated on 240 faulty Java program versions and compared with Ochiai, coverage representation learning, graph-based coverage localization, and variable-aware localization. The proposed method achieved 37.5% Top-1 accuracy and 91.7% Top-20 accuracy, while reducing the mean first rank to 6.8 statements. These results show that selective dependency analysis can improve localization precision and reduce developer inspection effort in complex software debugging tasks without creating excessive automated analysis overhead during diagnosis.