SBFL fault localization considering fault-proneness

Abstract

Fault localization is a critical phase in software debugging, often posing significant challenges and demanding extensive time for large and complex programs. Spectrum-based fault localization (SBFL) is a straightforward and cost-effective technique that leverages program execution logs to identify faulty statements. However, the effectiveness of SBFL can be compromised by biases in the test data set, which may not uniformly cover all code features. This study demonstrates that the integration of fault-proneness scores of program classes, predicted by a machine learning model utilizing source code metrics, with the fault-suspiciousness scores of program statements, estimated by SBFL, can enhance the accuracy and efficacy of fault localization. A Random Forest model is employed to predict the fault-proneness of classes in five Java projects from the Unified-Bug-Dataset 1.2. Concurrently, three established SBFL formulas are used to compute the fault-suspiciousness of statements. Statements are ranked based on their faultiness scores, derived from a linear combination of class fault-proneness and statement fault-suspiciousness. This approach is compared with the original SBFL formulas using four evaluation metrics: F-measure, precision, recall, and accuracy. The results indicate that the proposed method surpasses the original SBFL formulas across all metrics and significantly reduces the search space for fault localization. These findings suggest that the integration of static and dynamic analysis provides a more reliable and efficient method for fault localization in software systems.