A Framework for the Discovery of Predictive Fix-time Models
Francesco Folino, Massimo Guarascio, Luigi Pontieri · 2014
Fix-time prediction is a key task in bug tracking systems, which has been attracting the attention of data-mining researchers. However, traditional approaches only try to estimate the overall fix-time of a bug at the moment when it is initially reported, and do not care of updating such a preliminary estimate during the life of the bug. And yet, the sequence of actions performed on bug properties (e.g., its priority/criticality level, status or assignee) can help refine the prediction of the (remaining) fix-time for any bug still open, by adopting the strategy of predictive Process Mining techniques. Since typical bug-tracking systems lack of an articulated task-oriented description of the resolution process, we try to support the analyst in extracting relevant activities, and in mapping them to the low-level bug attribute modifications stored in bug histories. On the other hand, it was shown that the accuracy of fix-time predictors can be improved with the injection of derived data, encoding summarized statistics or high-level information. In order to address the above issues in a systematic manner, a comprehensive methodological framework is defined for the analysis of bug-like repositories, combining two kinds of tools: (i) modular and flexible data-transformation mechanisms, allowing the analyst to easily produce an enhanced process-oriented log, and (ii) a collection of induction techniques, for extracting effective prediction models from such a log. A toolkit supporting the whole approach was developed, and tested on real data from the bug repository of a popular open-source project. Preliminary results confirm the validity of our proposal, and the advantage of exploiting derived data in the prediction of fix times.