Classifying changes to models via changeset metrics

Automated classification of software changes can help to understand the reason why a change was made, guide the adoption of quality control practices as bugfix trends are observed, and cluster related sets of changes for similar management of the changed artifacts, thereby reducing maintenance efforts. A number of change classification techniques have been developed based on information extracted from the change author, change message, change size, or changed file. However, most of these approaches have targeted textual general-purpose programming languages. Furthermore, some of these approaches are computationally expensive because they often require the analysis of the whole source code, while others rely on the developers' ability to describe a commit via a well-written message. In this paper, we present an approach to classify changes to models into the appropriate maintenance type via a set of metrics that are extracted from the version history of models. We developed seven metrics related to changes applied to models and model elements. We then conducted an empirical study involving 10 classifiers to determine the classifier that offers the best performance for automating the change classification process. These classifiers were trained on over 300 changesets extracted from the version history of 28 Simulink repositories. The results of the study show that the Random Forest classifier offers the best performance. The Random Forest classifier has also been evaluated by comparing its results with labels extracted from the discussions within the issues reported in a similar time frame. The evaluation results show that the Random Forest classifier is able to achieve an F-1 score of 0.83, thereby showing its ability to classify changes into the appropriate categories intended by the original developers.