The nets I discovered for the Process Discovery Contest (PDC) 2019), classified 898 out of 900 correctly as fitting or non-fitting. Hence, I achieved with them a classification accuracy of 99.78%. Below, you’ll find all discovered nets.
All nets are sound WF-nets. Both nets 9 and 10 classify a single trace wrongly. All nets were discovered by using the “Log Skeleton Filter and Browser” to gain insights into the training event logs. Based on these insights, these nets were created manually, and checked against the training event logs using the standard replay techniques available in ProM.
For the nets, I used the “Apply Block Layout to Petri Net, User Selects Parameters” to obtain an initial layout. I then updated the layout manually where I deemed it necessary.
The classification results for model 7 (Net 7) came as a surprise. For the first and second intermediate test, Net 7 classified 4 and 5 traces (each out of 40) wrongly. For the final test, it classified all 90 traces correctly, which is really unexpected.
A ZIP archive containing the PNML files for the 10 nets can be downloaded here.
For the pictures above, I used a changed version of ProM that used a 30-points Arial font for the node labels. A regular version of ProM will use a much smaller font.