The Automated Discovery of Hybrid Processes Fabrizio M. Maggi University of Tartu Tijs Slaats* IT University of Copenhagen Exformatics Hajo A. Reijers VU University of Amsterdam
Overview Hybrid Process Models Discovering Hybrid Process Models Evaluation Future Work + Conclusion
Imperative Process Models
Flow-oriented Well-suited to rigid processes In a model with no flow nothing can happen Adding flow allows for additional possible behaviors Common in academia and industry
Declarative Process Models
Constraint-oriented Well-suited to flexible processes In an unconstrained model anything can happen Adding constraints limits behavior Still a novelty in industry
Hybrid Process Models
Different parts of the same process may be more or less flexible. Modeling a flexible process imperatively, or a strict process declaratively, often leads to incomprehensible models. Mixing of paradigms on the sub-process level: – Pockets of flexibility in workflow services [Sadiq et al.] – Flexibility as a Service (FAAS) [Aalst et al.]
Process Discovery Event Log Process Model
Process Discovery Current discovery techniques: – Mining Petri-nets / Flowcharts Alpha miner, Heuristic Miner, ILP miner, … – Mining Declarative constraints Declare miner But what if the log contains both flexible and rigid parts? – Imperative miners tend to blow-up on flexible logs – Declarative miners will need to find many constraints to model the strict parts of the process and will often have trouble finding all of them (resulting in processes with low precision) Solution: Hybrid Process Discovery!
Hybrid Process Discovery Context analysis Clustering (based on context analysis) Clustering (association rule mining) Standard Process Discovery Declare Discovery String Edit Distance
Evaluation – BPI Challenge 2012 Results of Imperative Miners
Evaluation – BPI Challenge 2012 Result of Hybrid Miner
Evaluation – BPI Challenge 2012 Comparison of Results Fitness Size
Future Work Proper plugin for Prom. Visualization of resulting hybrid model. Further evaluation on real cases. Further refinement of the heuristics used in the approach, for example the thresholds used for determining if an event is structured or unstructured.
Conclusion We offer the first automated approach for discovering hybrid process models. Using the approach on existing logs gives encouraging results: in particular for semi- structured logs the discovered models become more readable. Plenty of room for future work in an exciting new angle on process mining.