CamundaCon 2018: Profile Print and Explorative Data Mining (Duni, Viadee)
1 like593 views
This document discusses using machine learning to analyze process duration data from a process management system. It summarizes the following: - They analyzed process duration data from new orders to production ready to predict individual process instance durations and optimize scheduling and production planning. - Their machine learning model used gradient boosting and was trained on 75% of process instance data and validated on 25% to predict durations of all cases. - They explored using explanation models to understand what factors influence predicted durations and ensure the model can be trusted. - Integration scenarios were discussed to push predictions to services or batch predict, with the goal of a reusable and configurable pipeline.
CamundaCon 2018: Profile Print and Explorative Data Mining (Duni, Viadee)
- 1. Profile Print and Explorative Data Mining Fritz Ulrich (Duni) Frank Köhne (viadee)
- 2. 2
- 3. 3
- 4. 4
- 5. 5 Our process and how it should run Return order New order
- 6. 6 How does it look for Users?
- 7. 726.09.2018 © viadee 2018 Are we following the happy path?
- 8. 8 Now we can have a look at reports…
- 9. 9 …but what about the future?
- 10. 1026.09.2018 © viadee 2018 Objectives • Predicting the duration of an individual process instance • Optimized scheduling • Optimized production planning • Explain the factors that influence process duration • Explore what is possible
- 11. 1126.09.2018 © viadee 2018 DATA ANALYSIS • Mean duration: • New order production ready • Including both design, QC and customer approval! Duration bias! - Early in the process no long running process ended - Long running processes may be running still - Short ones are overrepresented
- 12. 1226.09.2018 © viadee 2018 Training of the model Validation of the model Create machine learning model 75% Prediction of all cases 986 Instances 25% 3.389 process instances (partly running)
- 13. 1326.09.2018 © viadee 2018 Prediction method • Gradient Boosting • State of the Art - Procedure for regression analyses • Compared to other possible models, fast, simple and high accuracy • How does the prediction work? 1. A simple model is trained. 2. Prediction of the model is compared with the real data. 3. A new model is trained to learn the errors of the old model and to correct them. 4. The two models are combined and step 2 is continued until the desired accuracy or maximum complexity is reached.
- 14. 1426.09.2018 © viadee 2018
- 15. 1526.09.2018 © viadee 2018 AI Explanation models 1. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Why should i trust you?: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1135–1144. ACM, 2016. 2. Ribeiro, M. T., Singh, S., & Guestrin, C. (2018). Anchors: High-Precision Model-Agnostic Explanations. Aaai. 1 2
- 16. 1626.09.2018 © viadee 2018 ORDER COMMENT LENGTH
- 17. 17 MACHINE LEARNING PIPELINE 26.09.2018 © viadee 2018 HistoryEventListener - Catch Events like a history table would. - Asynchronous and lightweight. Event Store / Stream - Store Events persistently, scaleable - Unchanged single point of truth Preprocessing (1/2) - Filter and flatten data - Cleansing - Aggregation to a data mining table for learning - Joins, anonymisation Learning - Data Mining - Visualisation - Explanation - Provide executable model (java) Reusable / Configurable Code Specific Code
- 18. INTEGRATION SCENARIOS HistoryEventListener Preprocessing Machine Learning B) Prediction Service Interface (Pull) A) Predict-on- Event (Push) C) Batch- Prediction (Push)
- 19. 1926.09.2018 © viadee 2018 LESSONS LEARNED • „Small Data“ – Meaningful analysis on limited data at process instance level • you may not need activity instance data at all • Camunda is useful both as a datasource, AI integration style and as a means to organize the lifecycle of prediction models. • Need for a governance process (new variables, renamed variables, new categorial values). • Discussion needed? • There may be considerable value in understanding the prediction model – even if it is not (yet) integrated in an automated process.
- 20. Thank you for your attention