Automated Bug Report Field Reassignment and Refinement Prediction
IEEE Transactions on Reliability
Bug fixing is one of the most important activities in software development and maintenance. Bugs are reported, recorded, and managed in bug tracking systems such as Bugzilla. In general, a bug report contains many fields, such as product, component, severity, priority, fixer, operating system (OS), platform, etc., which provide important information for the bug triaging and fixing process. Our previous study finds that approximately 80% of bug reports have their fields reassigned and refined at
... least once, and bugs with reassigned and refined fields take more time to fix than bugs with no reassigned and refined fields. Thus, automatically predicting which bug report fields get reassigned and refined could help developers to save bug fixing time. Considering that a bug report could have multiple field reassignments and refinements (e.g., the product, component, fixer, and other fields of a bug report can get reassigned and refined), in this paper, we propose a multi-label learning algorithm to predict which bug report fields might be reassigned and refined. We note that the number of bug reports with some types of reassignment and refinement (e.g., bugs whose severity fields gets reassigned and refined) is a small proportion of the whole bug report collection, indicating the class imbalance problem. Thus, we propose imbalanced ML.KNN (Im-ML.KNN), which extends ML.KNN, one of the state-of-the-art multi-label learning algorithms, to achieve better performance. Im-ML.KNN is a composite model that combines 3 multi-label classifiers built using different types of features (i.e., meta, textual, and mixed features). We evaluate our solution on 4 large bug report datasets including OpenOffice, Netbeans, Eclipse, and Mozilla containing a total of 190,558 bug reports. We show that Im-ML.KNN can achieve an average F-measure score of 0.56-0.62. We also compare Im-ML.KNN with other state-of-art methods, such as the method proposed by Lamkanfi et al., ML.KNN, and HOMER-NB. The results show that Im-ML.KNN, on average, improves the average F-measure scores of Lamkanfi et al.'s method, ML.KNN, and HOMER-NB by 119.69%, 9.11%, and 161.08%, respectively.