- Training log (XLog)
- Test log (XLog)
- Test log (XLog) with classified traces (using DrFurby Extension, see below).
DrFurby Classifier Plug-in
The DrFurby Classifier plug-in takes a training log and a test log, and returns a copy of the test log where every trace in the test log is classified as fitting (positive) or non-fitting (negative) to the training log. The main idea behind this plug-in is that weuse different techniques to avoid false positives and false negatives as much as possible:
- False negatives are avoided as much as possible by only using techniques that guarantee perfect fitness. As a result, none of the traces from the training log will be classified as negative.
- False positives are avoided as much as possible by using a knock-out system among these techniques. If one technique considers the trace to be negative, it will be negative.
At the moment, the following techniques are used:
- The Inductive Miner – infrequent with the noise setting set to 0.0 (which guarantees perfect fitness), combined with maximal decomposition (called IM-100).
- The Hybrid ILP Miner, combined with no decomposition (called HIM-0).
If too many false positives remain, then other perfect-fitness guaranteeing techniques (ILP Miner, non-maximal decomposition, …) could be added to the plug-in, leading to additional classifications per trace. So far, it seems to work okay, as adding these other techniques seems not to improve the classification.
A specific DrFurby Extension is used to enrich the copy of the test log with classification data. This extension provides the following attributes:
|Classification of the trace (“positive” or “negative”)
|The costs of replaying this trace on the net as discovered by HIM-0
|The costs of replaying this trace on the net as discovered by IM-100
|The number of milliseconds to it took classify the log
|The name of the classified log
|The number of traces classified negative
|The number of traces classified positive
|The accumulated costs of replaying this trace on all discovered nets
Our “DrFurby” submission to the contest can be found here.