H. M. W. Verbeek and F. Mannhardt, The DrFurby Classifier submission to the Process Discovery Contest @ BPM 2016, BPMCenter.org, BPM Center Report BPM-16-08, 2016.
To check the current state-of-the-art in process discovery, a contest has been set up. In this contest, 10 undisclosed process models were created by the organizers. For every created process model, the organizers created two logs: a training log containing 1000 traces and a test log containing 20 traces. Both logs were disclosed to the participants. The challenge for the participants was to classify every trace from every test log whether or not it ts the corresponding original process model, that is, whether or not it can be replayed perfectly by that process model. As the original process model was not disclosed, the participants had to first discover an as-good-as-possible process model from the training log, and second use this discovered process model to classify every trace from the corresponding test log. The main assumption here is that a better discovery technique discovers a better (that is, more matching to the original) process model, which then results in a better classification. In the end, the one that classifies the most traces correctly, wins the contest. This paper presents a submission to this contest that, in the end, classifies 193 out of 200 traces correctly.