Predicting Effectiveness of IR-Based Bug Localization Techniques

摘要: Recently, many information retrieval (IR) based bug localization approaches have been proposed in the literature. These use techniques to process a textual report and collection of source code files find buggy files. They output ranked list sorted by their likelihood contain bug. Recent can achieve reasonable accuracy, however, even state-of-the-art tool outputs lists where appear very low lists. This potentially causes developers distrust tools. Parnin Orso recently conduct user study highlight that do not an automated debugging useful if they root cause early list. To address this problem, we build oracle automatically predict whether produced IR-based is likely be effective or not. We consider file appears top-N position If unlikely effective, need waste time checking recommended one one. In such cases, it better for traditional methods request further localize bugs. oracle, our approach extracts features divided into four categories: score features, topic model metadata features. separate prediction each category, combine them create composite which used as oracle. name APRILE, stands Automated Prediction Bug Localization's Effectiveness. evaluated APRILE effectiveness three IR tools on more than thousands reports from AspectJ, Eclipse, SWT. average precision, recall, F-measure at least 70.36%, 66.94%, 68.03%, respectively. Furthermore, outperforms baseline 84.48%, 17.74%, 31.56% SWT reports,