An Approach to the Discovery of Accurate and Expressive Fix-Time Prediction Models

Abstract

Predicting the fix time (i.e. the time needed to eventually solve a case) is a key task in an issue tracking system, which attracted the attention of data-mining researchers in recent years. Traditional approaches only try to forecast the overall fix time of a case when it is reported, without updating this preliminary estimate as long as the case evolves. Clearly, the actions performed on a case can help refine the prediction of its (remaining) fix time, by using Process Mining techniques, but typical issue tracking systems lack task-oriented descriptions of the resolution process, and store fine-grain records, just registering case attributes’ updates. Moreover, no general approach has been proposed in the literature that fully supports the definition of high-quality derived data, which were yet proven capable to improve prediction accuracy considerably. A new fix-time prediction framework is presented here, along with an associated system, both based on the combination of two kinds of capabilities: (i) a series of modular and flexible data-transformation mechanisms, for producing an enhanced process-oriented log view, and (ii) several induction techniques, for extracting a prediction model from such a view. Preliminary results, performed on the logs of two real issue tracking scenarios, confirm the validity and practical usefulness of our proposal.