Predictive Reranking using Code Smells for Information Retrieval Fault Localization

Code metrics and static checker information can be used to increase the performance of Information Retrieval Fault Localization (IRFL). While previous work focused on the calculation of bug proneness scores for reranking IRFL tool outputs, we propose a dynamic reranking approach based on code smells. We hypothesize that code smell information can be predicted from textual bug reports and that these insights can support the fault localization task. To test this hypothesis, we implemented a fault localization pipeline consisting of a static checker (PMD) and a Neural Network based classifier to predict code smells. We apply this pipeline to predict code smells for a given textual bug report and use this information to rerank the outputs of BugLocator, BLIA, and BRTracer. We trained and evaluated our pipeline on the Bench4BL bug benchmark. While overall performance on the Bench4BL benchmark stayed the same, we found significant MAP localization performance increases for specific input software projects. We identify and discuss the characteristics of debugging benchmarks that limit their suitability for machine learning experiments. Despite these issues, we demonstrated that predicting certain smells based on textual bug reports is possible and that this information may be beneficial to fault localization.