International Conference on Business Process Management 2018 (BPM2018 Sydney).
Performance is central to processes management and event data pro-vides the most objective source for analyzing and improving performance. Currentprocess mining techniques give only limited insights into performance by aggre-gating all event data for each process step. In this paper, we investigate processperformance of all process behaviors without prior aggregation. We propose theperformance spectrumas a simple model that maps all observed flows betweentwo process steps together regarding their performance over time. Visualizing theperformance spectrum of event logs reveals a large variety of very distinctpatternsof process performanceand performance variability that have not been describedbefore. We provide a taxonomy for these patterns and a comprehensive overviewof elementary and composite performance patterns observed on several real-lifeevent logs from business processes and logistics. We report on a case study whereperformance patterns were central to identify systemic, but not globally visibleprocess problems.
3. Example of BHS Performance Problem
2
Check-In
Transfer
Scanner 1
Scanner 2
Merge
Sorter
Exit
Dead-lock on the Sorter Loop
4. Example of a Real BHS Model
3
Transfer
Scanner 2
Sorter
Exit
5. Questions about the model:
• Do all cases take 12m from Scanner 2 to Merge?
• When and why did bottlenecks happen?
• How do process variants divert/merge?
BHS: Questions about Performance
4
Check-In Transfer
Merge
Sorter
Scanner 1 Scanner 2
Exit
1m 1m
10m 12m
6m
3m
Baggage Handling
System
6. Questions about the model:
• Do all cases take 12m from Scanner 2 to Merge?
• When and why did bottlenecks happen?
• How do process variants divert/merge?
BHS: Questions about Performance
5
Check-In Transfer
Merge
Sorter
Scanner 1 Scanner 2
Exit
1m 1m
10m 12m
6m
3m
Baggage Handling
System
Generalized questions:
• What is the actual distribution of durations?
• What are timelines of individual cases?
• What is performance over time?
• How do cases and process steps influence each
other?
7. Business Process: Questions about Performance
6
Road Traffic Fines
Management process
Generalized questions:
• What is the actual distribution of durations?
• What are timelines of individual cases?
• What is performance over time?
• How do cases and process steps influence each
other?
8. 1. Visual Analytics: a different way to visualize event data
2. New findings in well-known data
3. Introduction to performance patterns
4. Evaluation and a new tool
How to Represent System Performance in Dynamic?
7
Event Log Annotated Model
?
Performance Abstraction Levels
9. Performance in Dynamic
8
A
B
Sub-trace: A(time=t1) → B(time=t2)
Cases started/stopped/pending/…
normal
speed
2 times
slower
3 times
slower
very
slow
Time
A
B
t1 t2
StepStep
Performance Spectrum
10. 1. Visual Analytics: a different way to visualize event data
2. New findings in well-known data (Road Traffic Fines
Management process)
3. Introduction to performance patterns
4. Evaluation and a new tool
How to Represent System Performance in Dynamic?
9
Performance
Spectrum
Event Log Annotated ModelPerformance Abstraction Levels
13. Example: Complete Case Variant
12
Send for Credit
Collection
5d
Create Fine
Send Fine
Insert Fine
Notification
Add Penalty
14. 1. Visual Analytics: a different way to visualize event data
2. New findings in well-known data
3. Introduction to performance patterns
4. Evaluation and a new tool
How to Represent System Performance in Dynamic?
13
Performance
Spectrum ?
Event Log Annotated ModelPerformance Abstraction Levels
16. Full Taxonomy → See in Paper
15
ORDER
unordered FIFO LIFO
batching
on start
batching
on end
constant
speed
batching
on start and end
CLASSES PRESENTED
1 >1
WORKLOAD TRENDS
(TOTAL OR FOR A CLASS)
steady variable
growing falling
AMOUNT OF
WORKLOAD
zero non-zero
low medium high
SIZE
segment one
subsequence
ab
bc
cd
several
subsequences
kl
lm
mn
ab
bc
R1
R2
OCCURRENCE
globally
as a local
instance
REPETITIONS
single periodicsystematic
arbitrary
TYPE
detailed composite aggregated
VARIANTS
CONTAINED
1 >1
Aggregation functions: {cases started, cases stopped, diff. of start/stop, cases pending}
<exact number>
exact
number
spiky
scheduled
variable
speed
WORKLOAD CHARACTER
continuous sparse
PERFORMANCE IN
CONTEXT
slower faster
the same
WORKLOAD
PERFORMANCE
SCOPE
Classifiers:
quartile-based, median-proportional (for MHS)
DURATION= <abs. value>
SHAPE
17. 1. Visual Analytics: a different way to visualize event data
2. New findings in well-known data
3. Introduction to performance patterns
4. New tool and evaluation
How to Represent System Performance in Dynamic?
16
Event Log Annotated ModelPerformance Abstraction Levels
Performance
Spectrum
Performance
Patterns
18. Performance Spectrum Miner and Evaluation
17
https://github.com/processmining-in-logistics/psm
See it in action:
Demo Session 2!
✓ ProM plugin & Stand-alone
✓ Documentation
✓ Open-Source LGPL-3
✓ Java-compatible
• Real-world problem (Vanderlande, BHS of a major European airport)
• BPI Challenge 2018: applying to business processes
19. Business Process vs. Material Handling System
18
Road Traffic Fine Management Process Baggage Handling System
20. Spectra of Public Event Logs from 4TU Datacenter
19
BPI’12
BPI’14
BPI’15-1 BPI’15-2 BPI’15-5
BPI’17 Hospital Billing RTFM
21. New visual analytics technique:
1. performance of the whole process over time (all cases)
2. explore how process steps/cases influence each other
3. analyze processes on a higher abstraction level: performance patterns
Conclusions
20
Performance
Spectrum
Performance
Patterns
Event Log Annotated ModelPerformance Abstraction Levels
22. Future Work
21
• automatic ordering of segments
• dealing with parallel activities
• automatic detection of patterns and annotating process models with them
• predicting performance spectra to detect performance issues of MHS