Enabling Precise Interpretations of Software Change Data

Numerous techniques mine change data captured in software archives to assist software engineering efforts. These change-based approaches typically analyze change sets – groups of co-committed changes – under the assumption that the development work represented by change sets is both meaningful and related to a single change task. However, we have found that change sets often violate this assumption by containing changes that we consider to be non-essential, or less likely to be representative of the kind of meaningful software development effort that is most interesting to typical change-based approaches. Furthermore, we have found many change sets addressing multiple subtasks – groups of isolated changes that are related to each other, but not to other changes within a change set. Information mined from such change sets has the potential for interfering with the analyses of various change-based approaches. We propose a catalog of non-essential changes and describe an automated technique for detecting such changes within version histories. We used our technique to conduct an empirical investigation of over 30 000 change sets capturing over 25 years of cumulative development activity in ten open-source Java systems. Our investigation found that between 3% and 26% of all modified code lines and between 2% and 16% of all method updates consisted entirely of non-essential modifications. We further found that eliminating such modifications reduces the amount of false positive recommendations that would be made by an existing association rule miner. These findings are supported by a manual evaluation of our detection technique, in which we found that our technique falsely identifies non-essential method updates in only 0.2% of all cases. These observations should be kept in mind when interpreting insights derived from version repositories. We also propose a formal definition of “subtasks” and present an automated technique