Leveraging Evidence Theory to Improve Fault Localization: An Exploratory Study

Background: Fault localization in software maintenance and debugging can be a costly process. Spectrum-Based Fault Localization (SBFL) is a widely-used method for fault localization. It assigns suspicion scores to code elements based on tests, indicating the likelihood of defects in specific code lines. However, the effectiveness of SBFL approaches varies depending on the subject code. Aims: In this paper, our aim is to present an approach that combines multiple SBFL formulae using evidence theory. Method: We first introduce a taxonomy of SBFL techniques. Then, we describe how we fuse suspiciousness scores obtained from a set of SBFL formulae. We also introduce a concept of fuzzy windows, and describe how they can enhance localization accuracy and how they can be tuned to further refine results. Results: We present an empirical evaluation of our approach using the Defects4J dataset. Our results demonstrate improvements in fault localization accuracy over existing statement-level SBFL techniques. Specifically, by fusing three SBFL methods, our approach reduces code inspection effort by up to 34.5 % with a size-4 window and increases the hit rate for the top 10% most suspicious lines by 27.9 % using a size-7 window. Moreover, in multi-line bug scenarios, our approach reduces code inspection effort by up to 35.6% and achieves a maximum increase of 43.2% in the hit rate of the top 10% most suspicious lines. Additionally, our approach outperforms state-of-the-art machine learning-based method-level fusion approaches in terms of top rank fault localization accuracy. Conclusions: Our study highlights the applicability of evidence theory in addressing fault localization as an uncertain and ambiguous information fusion problem involving multiple SBFL techniques. The combination of SBFL formulae using evidence theory, along with the use of fuzzy windows, shows promise in enhancing fault localization accuracy.