Minimizing Overprocessing Waste in Business Processes via Predictive Activity Ordering
-
Upload
marlon-dumas -
Category
Education
-
view
254 -
download
0
Transcript of Minimizing Overprocessing Waste in Business Processes via Predictive Activity Ordering
Minimizing Overprocessing Waste in Business Processes via
Predictive Activity Ordering
Ilya Verenich, Marlon Dumas, Marcello La Rosa, Fabrizio Maggi, Chiara Di Francescomarino
Presentation at CAiSE’2016 – Ljubljana, 15 June 2016
2
Knockout section • One activity with a negative outcome “knocks-out” the case
• To avoid overprocessing, we should execute first the activity that will knock-out the case (if we knew it!)
3
Minimizing overprocessing waste Execute highly selective tasks first.
Execute tasks that raise problems first Postpone expensive tasks until the end
Design-time approach (Aalst 2001) Our approach
Order checks by probability of case rejection and mean effort
• Reject probabilities and effort and constant for each case
• Does not take into account specifics of each case
• These values are specific for each case
• They are estimated via predictive models
4
Processing effort and overprocessing waste
• Minimum processing effort:
• (actual) Processing effort:
• Overprocessing:
How can we know the actual processing effort?
5
Expected processing effort
• Knockout section with three activities:
• Reject probability of an activity
6
Expected processing effort (cont’d)
• Knockout section with three activities:
• Knockout section with N activities:
7
Our approach
8
Our approach
9
Our approach
10
Our approach
13
Datasets
Bondora online P2P lending:• > 45K process cases• Knockout section with 3 independent activities,
P=(0.08,0.03,0.05)• > 30 case attributes
Environmental permit log (CoSeLoG project):• ca 1400 process cases• Knockout section with 3 semi-independent activities,
P=(0.01,0.01,0.61)• 4 case + 2 event attributes
14
Evaluation of predictive models: ROC
15
Number of checks required
• 1, if there will be at least one activity that will reject the case OR
• 3, otherwise
16
Evaluation – reduction in # of checks
Avg # of checks reduced with our approach
Overprocessing is reduced
17
Conclusion
• Using predictive models reduces overprocessing• Performance depends on the difference between average
rejection rate of checks• More experiments are needed for real-world scenarios (checks
can be dependent, etc.)
18
Thank you
Q&A