From Low-Level Events to Activities - A Pattern-based Approach
1 like145 views
The document presents a pattern-based approach for abstracting low-level events into high-level activities in process mining. It discusses the challenges of event recognition and outlines a method that incorporates knowledge of activity patterns and alignment techniques to improve abstraction reliability. The proposed approach has been evaluated in a hospital information system, demonstrating its effectiveness in handling shared labels, concurrency, and noise.
![3) Align event log to abstraction model
Low-level Event Time
Red Button 20:08:00
Green Button 20:10:00
Green Button 22:02:00
Nurse 22:09:00
Gray Button 22:10:00
Nurse 22:15:00
Nurse 07:02:00
Green Button 07:15:00
Gray Button 07:18:00
Process Step Time
RedA 20:08:00
GreenA 20:10:00
GrayA
GreenV 22:02:00
GrayV 22:10:00
NurseS 07:02:00
GreenS 07:15:00
GrayS 07:18:00
Event Log Existing alignment methods [1] Aligned Log
Model only
Log only
…
[1] Balanced multi-perspective checking of process conformance.
Computing. 98 (4). 2016](https://image.slidesharecdn.com/emisa-eventabstraction-170614085906/85/From-Low-Level-Events-to-Activities-A-Pattern-based-Approach-14-320.jpg)
From Low-Level Events to Activities - A Pattern-based Approach
- 1. From Low-Level Events to Activities A Pattern-based Approach Felix Mannhardt, Massimiliano de Leoni, Hajo A. Reijers, Wil M.P. van der Aalst, Pieter J. Toussaint (NTNU, Trondheim) Presented at BPM 2016: 10.1007/978-3-319-45348-4_8
- 2. Problem: Events ≠ Recognizable high-level activities PAGE 1 Alarm records Event Time Red Button 20:08:00 Green Button 20:10:00 Gray Button 20:16:00 Shift Event Time Nurse 07:02:00 Green Button 07:15:00 Gray Button 07:18:00 records Event Time Green Button 22:02:00 Gray Button 22:10:00 records Visit
- 3. Goal 1: From low-level events to high-level activities PAGE 2 Event Log Abstracted Log Low-level Event Time Red Button 20:08:00 Green Button 20:10:00 Gray Button 20:16:00 Green Button 22:02:00 Gray Button 22:10:00 Nurse 07:02:00 Green Button 07:15:00 Gray Button 07:18:00 High-level Event Start Complete Alarm 20:08:00 20:16:00 Visit 22:02:00 22:10:00 Shift 07:02:00 07:18:00
- 4. Goal 2: Deal with shared labels, concurrency and noise PAGE 3 Low-level Event Time Red Button 20:08:00 Green Button 20:10:00 Gray Button 20:16:00 Green Button 22:02:00 Nurse Changed 22:09:00 Gray Button 22:10:00 Nurse 22:15:00 Nurse 07:02:00 Green Button 07:15:00 Gray Button 07:18:00 High-level Activity Start Complete Alarm 20:08:00 20:10:00 Visit 22:02:00 22:10:00 Shift 07:02:00 07:18:00 Missing events Unexpected events Event Log Abstracted Log
- 5. Related work • Unsupervised event abstraction (Ferreira et al., Folino et al., …) • Does not take domain knowledge into account • Supervised event abstraction (Thomas Baier et al., Niek Tax, …) • Assumes knowledge of a complete process model • Semi-automatic discovery of the mapping between events and activities • Uses clustering and constraint satisfaction to determine the mapping • Uses annotated event logs • Complex event processing • Focus on detection of event patterns in data streams • Not directly considering process instances / traces PAGE 4
- 6. Overview: From Low-level Events to Activities PAGE 5 Event Log Aligned Log Abstracted Log Activity Pattern Abstraction Model 3) Align model & log 4) Abstract events 1) Encode knowledge 2) Compose model
- 7. 1) Encode knowledge on activities as activity patterns PAGE 6 Alarm Shift GreenV GrayV Pattern defines traces expected for one activity instance! Data Petri net used for clear semantics! Visit
- 8. What about interaction between activity patterns? PAGE 7 Alarm Interaction? GreenV GrayV Visit
- 9. Overview – Step 2 – Composition PAGE 8 Event Log Aligned Log Abstracted Log Activity Pattern Abstraction Model 2) Compose model
- 10. 2) Build an integrated abstraction model PAGE 9 Interaction? Alarm Visit Parallel Alarm Visit Interleaving ↔ Alarm Visit Choice ✖ Alarm Visit Sequence Alarm Visit Alarm 0..✱ Repetition
- 11. 2) Build an integrated abstraction model PAGE 10 Handover Alarm Visit ↔ 0..✱ 0..✱ 0..✱ 0..✱ Compile Abstraction Model Compiled Abstraction Model (DPN)
- 12. Overview – Step 3 – Alignment PAGE 11 Event Log Aligned Log Abstracted Log Activity Pattern Abstraction Model 3) Align model & log
- 13. 3) Align event log to abstraction model PAGE 12 Event Log Low-level Event Time Red Button 20:08:00 Green Button 20:10:00 Green Button 22:02:00 Nurse 22:09:00 Gray Button 22:10:00 Nurse 22:15:00 Nurse 07:02:00 Green Button 07:15:00 Gray Button 07:18:00 Compiled abstraction model Alignments
- 14. 3) Align event log to abstraction model Low-level Event Time Red Button 20:08:00 Green Button 20:10:00 Green Button 22:02:00 Nurse 22:09:00 Gray Button 22:10:00 Nurse 22:15:00 Nurse 07:02:00 Green Button 07:15:00 Gray Button 07:18:00 Process Step Time RedA 20:08:00 GreenA 20:10:00 GrayA GreenV 22:02:00 GrayV 22:10:00 NurseS 07:02:00 GreenS 07:15:00 GrayS 07:18:00 Event Log Existing alignment methods [1] Aligned Log Model only Log only … [1] Balanced multi-perspective checking of process conformance. Computing. 98 (4). 2016
- 15. Overview – Step 4 – Abstract Events PAGE 14 Event Log Aligned Log Abstracted Log Activity Pattern Abstraction Model 4) Abstract events
- 16. 4) Create an abstracted event log for high-level activities PAGE 15 Aligned Log Abstract aligned events Abstracted Log High-level Event Transition Time Alarm start 20:08:00 Alarm complete 20:10:00 Visit start 22:02:00 Visit complete 22:10:00 Shift start 07:02:00 Shift complete 07:18:00 Process Step Time RedA 20:08:00 GreenA 20:10:00 GrayA GreenV 22:02:00 GrayV 22:10:00 NurseS 07:02:00 GreenS 07:15:00 GrayS 07:18:00 Alarm Visit Shift Use time of previous mapped event. Alignment forced a model move: Matching error! Additional data attributes can be mapped!
- 17. Recap: Abstraction Method PAGE 16 Event Log Aligned Log Abstracted Log Activity Pattern Abstraction Model 3) Align event log to abstraction model 4) Abstract based on aligned log 1) Encode knowledge on activities as patterns 2) Build integrated model
- 18. Evaluation: Digital whiteboard system in a hospital PAGE 17 • Information system • Digital whiteboard • Supports work of nurses • Mixed clinical & logistic info • Flexible system • Dataset • One year • > 8,000 cases • > 280,000 events • Event per changed cell
- 19. Evaluation: Activity Patterns PAGE 18 • 18 activity patterns • Our assumptions • Interview with expert • Abstraction model • Most interleaved & repeated • Five concurrent activities • Resulting abstraction • Low average error rate • Successful abstraction
- 20. Evaluation: Detected shift change pattern PAGE 19 Shift pattern captures a meaningful activity Blue: Nurse Green: Call Signal Green Yellow: Call Signal Gray Relatively rare pattern! 00:00 24:00 00:00 24:00
- 21. Conclusion & Future work PAGE 20 • Method • Pattern-based event abstraction • Knowledge encoded as activity patterns • Abstraction using alignment methods • Results • Handles shared labels, concurrency and noise • Alignment gives reliability measure • Successfully used in case studies • Future work • Prioritization among activity patterns • Decomposed / approximate alignment methods • Mining / recommendation of suitable patterns Implemented in ProM 6.6
- 22. Questions? PAGE 21 @fmannhardt - f.mannhardt@tue.nl fmannhardt.de/g/pba - Documentation & Installation
- 23. Backup: Mapping of the interactions to DPN PAGE 22 Parallel Alarm Visit Interleaving ↔ Alarm Visit Choice ✖ Alarm Visit Sequence Alarm Visit Alarm Visit source sink Alarm Visit source sink Alarm Visit source sink Alarm Visit source sink
- 24. Backup: Result using IVM PAGE 23 Examinations recorded by multiple TreatmentChanged Unabstracted events Abstracted events 300 patients with Chest Pain receive an X-Ray Transfer Discharge
Editor's Notes
- #2: Good afternoon, I would like to present our work about “Pattern-based Event Abstraction”. My name is Felix Mannhardt, I’m a PhD student at the Eindhoven University of Technology. This paper is joint work with Massimiliano, Hajo, Wil from Eindhoven and Pieter Toussaint from NTNU, Trondheim.
- #3: Assume that you want to analyze work at a hospital ward For example, you want to look at three activities: Alarm (activated by a patient) Visit (by a nurse) Handover (from a nurse to another nurse) Looking at events in your event log only events at a lower level of abstraction are recorded For example, Alarm corresponds to: Red Button, Green Button, Gray Button … Our work had two goals
- #6: Related work (PhD thesis by Thomas Baier): Assume knowledge of the full process model for the overall process Resolves to clustering methods and heuristics for concurrency and noise
- #9: Can both activities occur concurrently, are there constraints on the cardinality, are they mutually-exclusive
- #12: Be quick here: For details see the paper & technical report!
- #17: Remember the other data
- #25: TODO: maybe show the nurse call model instead! TODO: fix the model for the unabstracted events