A fusion approach for classifying duplicate problem reports

Issue tracking systems play a critical role in software maintenance by allowing users and developers to submit problem reports for observed failures. A major problem in these systems is that two or more users can, and do, submit reports describing the same issue. Automated classification of such duplicate problem reports is an area of active research. The corpus of existing research shows a slow improvement in classification accuracy using relatively small subsets of problem report data. When applied to an entire project's problem repository, they exhibit a reduction in performance. In this paper we propose a novel duplicate report detection approach using multi-label classification. We use a suite of 24 duplicate classification techniques and MULAN software package to train a multi-label classifier. This multi-label classifier selects a set of similarity measures (from a pool of measures) that are most likely to find the true primary report. To demonstrate its effectiveness the method was tested on the entire Firefox repository. This data set encompasses 12+ years of problem reports and contains over 30,000 duplicate reports. Our results indicate that multi-label classification boosts the performance of the individual measures by up to 40% while returning overall results that match or outperform existing methods. The proposed method uses less than 1% of the dataset for training.