Skip to main content
Solved

What kind of algorithms does Process Analytics use for discovery and conformance?

  • August 9, 2021
  • 4 replies
  • 20 views
.
Level 6

As we know, Process Mining has the tremendous advantage of providing an objective and realistic view on the processes, as opposed to the traditional BPM analysis, which is very dependent on human judgement and prone to bias.

However, building a process model from an event log is not a straightforward task. It requires an algorithm, and algorithms are based on assumptions have to somehow deal with things like noise. Conformance checking can also be done in different ways.

 

Therefore, to better understand my results, I would be very interested in learning the algorithms used by Celonis. Thanks!

Best answer by joos.buijs

Hi @1460047871 ,

 

I hope you'll get an official reply from Celonis.

 

In the mean time, I have found that the process discovery is done by constructing a Transition System using directly follows relations (i.e. an arc indicates a directly follows relation). (there are papers, for instance by H.M.W. Verbeek on transition systems that provide more details).

Most important observation here is that not a real process is discovered, but only the directly follows relations are visualized.

 

As for conformance checking, an algorithm (based on) token based replay is used.

 

Fun fact: if you go into conformance checking, create a model, and choose for discovery, the inductive miner algorithm (by S.J.J. Leemans) is used to discover a BPMN model.

 

I hope this helps!

 

And if anyone has other insights I'm really looking forward to them!!!

4 replies

joos.buijs
Level 10
Forum|alt.badge.img+2
  • joos.buijs
  • Level 10
  • Answer
  • August 10, 2021

Hi @1460047871 ,

 

I hope you'll get an official reply from Celonis.

 

In the mean time, I have found that the process discovery is done by constructing a Transition System using directly follows relations (i.e. an arc indicates a directly follows relation). (there are papers, for instance by H.M.W. Verbeek on transition systems that provide more details).

Most important observation here is that not a real process is discovered, but only the directly follows relations are visualized.

 

As for conformance checking, an algorithm (based on) token based replay is used.

 

Fun fact: if you go into conformance checking, create a model, and choose for discovery, the inductive miner algorithm (by S.J.J. Leemans) is used to discover a BPMN model.

 

I hope this helps!

 

And if anyone has other insights I'm really looking forward to them!!!

.
Level 6
  • .146001171
  • Author
  • Level 6
  • August 10, 2021

Hey, @joos.buijs 

 

Thank you for your reply.

 

Indeed, I was guessing it was a Transition System. Therefore, I wonder what happens when the process to be discovered has concurrency (AND-splits)? Because theoretically, if there are n concurrent activities, the Transition System will involve up to n! variants, which adds unnecessary complexity.

 

As for conformance checking, I also noticed that the tool for creating a process model does not allow to insert AND gateways... Maybe concurrency is out of scope in this software.

janpeter.van.d
Level 12
Forum|alt.badge.img+26

Hey, @joos.buijs 

 

Thank you for your reply.

 

Indeed, I was guessing it was a Transition System. Therefore, I wonder what happens when the process to be discovered has concurrency (AND-splits)? Because theoretically, if there are n concurrent activities, the Transition System will involve up to n! variants, which adds unnecessary complexity.

 

As for conformance checking, I also noticed that the tool for creating a process model does not allow to insert AND gateways... Maybe concurrency is out of scope in this software.

Hi Rui,

 

Due to the lack of parallel flows, the model has indeed more (unneccessary) variations. In previous versions of Celonis, you were able to define parallel flows by having a parent and sub process column, see this post from the older community (unfortunately not all images are migrated well): https://www.celopeers.com/s/question/0D50700000GC4mdCAD/parallel-processes-subprocesses

 

To better show parallel processes, you could use the multi-eventlog, but this is not exactly the same functionality (i.e. you have to define the split, instead that Celonis finds it automatically) and requires a bit more work to set up.

 

Kind regards,

Jan-peter

.
Level 6
  • .146001171
  • Author
  • Level 6
  • August 11, 2021

Hi Rui,

 

Due to the lack of parallel flows, the model has indeed more (unneccessary) variations. In previous versions of Celonis, you were able to define parallel flows by having a parent and sub process column, see this post from the older community (unfortunately not all images are migrated well): https://www.celopeers.com/s/question/0D50700000GC4mdCAD/parallel-processes-subprocesses

 

To better show parallel processes, you could use the multi-eventlog, but this is not exactly the same functionality (i.e. you have to define the split, instead that Celonis finds it automatically) and requires a bit more work to set up.

 

Kind regards,

Jan-peter

Thank you very much @janpeter.van.d !

 

Community Code of ConductAccessibility statement