In brief
Discovery produces a model from an event log without prior information. Conformance compares an existing model with the event log of the same process. Enhancement extends or repairs a model using information from the data. Conformance measures the gap; enhancement changes the model.
Three types, three data situations
The Process Mining Manifesto of the IEEE Task Force on Process Mining distinguishes three basic types. The distinction is not academic — it determines what data a project needs in the first place.
| Basic type | Input | Output | Question |
|---|---|---|---|
| Discovery | Event log | Process model | How does the process actually run? |
| Conformance | Event log and model | Diagnostics of deviations | Does reality follow the target model? |
| Enhancement | Event log and model | Improved model | What do the data say beyond the model? |
Discovery takes an event log and produces a model without using any a-priori information. For many organisations this is the most surprising part, because real processes can be reconstructed from example executions alone.
Conformance compares an existing model with the event log of the same process — in both directions. The reference is not limited to procedural models: organisational models, business rules, policies and laws qualify as well.
Enhancement extends or repairs an existing model. Timestamps, for instance, let bottlenecks, service levels, throughput times and frequencies be written into the model.
The confusion that derails projects
Conformance measures the gap between model and reality. Enhancement changes the model. Anyone looking for deviations while adjusting the model ends up with no deviations — and nothing learned.
This rarely happens out of ignorance. It happens because a model that does not fit is uncomfortable and easy to adjust. The sequence therefore has to be fixed in the project: establish conformance first, then decide whether the model or reality was wrong.
Which type fits which question
- There is no reliable target model → discovery. The result is a description, not an assessment.
- There is a target model and the question is "is it being followed?" → conformance. The result is a diagnosis.
- There is a model but no figures attached to it → enhancement. The result is the same model with throughput times and bottlenecks.
The manifesto also makes three things explicit: process mining is not limited to control-flow discovery, it is not a special case of data mining — the authors call it the missing link between data mining and model-driven business process management — and it is not limited to offline analysis: the techniques work on historical data, but the results apply to running cases.
Four questions before scoping
- Does a target model exist — and is it current enough to serve as a reference?
- Should the result describe, assess or extend?
- Who is allowed to change the model when the data contradict it?
- Do the timestamps carry the meaning the analysis assumes?
Basis
van der Aalst, W. et al. (2012): Process Mining Manifesto, IEEE Task Force on Process Mining, LNBIP 99, pp. 169–194, Springer.