1. Decomposed Process Mining: The ILP Case
Eric Verbeek and Wil van der Aalst

2. A Problem
/ department of mathematics and computer science

3. A Solution
/ department of mathematics and computer science

4. Regular Discovery / department of mathematics and computer science

5. Decomposed Discovery: Divide
/ department of mathematics and computer science

6. Decomposed Discovery: Conquer
/ department of mathematics and computer science

7. Regular Replay / department of mathematics and computer science

8. Decomposed Replay: Divide
/ department of mathematics and computer science

9. Decomposed Replay: Conquer
/ department of mathematics and computer science

10. Example Model (Accepting Petri Net)
/ department of mathematics and computer science

11. Example Event (Activity) Log
/ department of mathematics and computer science

12. Divide and Conquer Framework
/ department of mathematics and computer science

13. Divide and Conquer Framework
See
/ department of mathematics and computer science

14. Decomposed ILP Discovery Algorithm
/ department of mathematics and computer science

15. Filter (First Cluster)
Filter In/Out
Replace
/ name of department

16. Decomposed ILP Discovery Algorithm
/ department of mathematics and computer science

17. Discovered Model / department of mathematics and computer science

18. Decomposed ILP Replay Algorithm
/ department of mathematics and computer science

19. Strategy (Third Cluster)
Filter
Replace
/ department of mathematics and computer science

20. Decomposed ILP Replay Algorithm
/ department of mathematics and computer science

21. Replay Cost Factor Implementation issue: Solution:
The ILP-based replayer takes integer costs.
If an activity occurs in, say, 3 clusters, then the replay costs of this activity in a single cluster should be a third of the usual replay costs in the entire model.
Solution:
Take the greatest common divisor of all activity cluster counts, multiply all replay costs by that factor, and later on divide all replay costs by this factor again.
/ department of mathematics and computer science

22. Decomposed ILP Replay Algorithm
/ department of mathematics and computer science

23. Case Study Setting Mode Event log based on BPI Challenge 2012 log
Model discovered in earlier work
Event log aligned on discovered model
/ department of mathematics and computer science

24. Case Study Model / department of mathematics and computer science

25. Case Study Results / department of mathematics and computer science

26. Conclusions General framework for decomposed process mining
Objects with imports, exports, and visualizers
Accepting Petri Net
Causal Activity Matrix
Causal Activity Graph
Activity Cluster Array
Event Log Array
Accepting Petri Net Array
Log Alignment
Log Alignment Array
Many algorithms
/ department of mathematics and computer science

27. Conclusions
ILP-based decomposed discovery and ILP-based decomposed replay
Discovery can result in the same model in a fraction of the time
Replay can result in less costs in less time (trade-off)
/ department of mathematics and computer science

28. Future Work Non-maximal decompositions
Grouping maximally-decomposed clusters may be beneficial (work of Bart Hompes)
Splitting large maximally-decomposed clusters may also be beneficial (cf. original BPI Challenge 2012 log)
Support for different discovery and replay algorithms
Merging nets
Merging alignments
/ department of mathematics and computer science