PAIS Lab seminar: Fast and Accurate Declare Mining // M. Westergaard

Monday, September 23^th, 18:30, aud. 902 (Kirpichnaya, 33)

Michael Westergaard,
researcher at TU/e (Eindhoven Technical University, The Netherlands),
leading research fellow of PAIS Lab (Higer School of Economics).

Michael obtained his PhD from Aarhus University in 2007. The topic was improvement of complex systems using colored Petri nets, model-checking, and behavioral visualization. Starting in 2010, he has been working at Eindhoven University of Technology with more dynamic modeling languages, in particular the declarative process description language Declare. The last year, Michael has had a shared position between Eindhoven and the Higher School of Economics, and has been working on combining the more imperativenature of colored Petri nets with Declare. Concrete outcomes of this area hybrid modeling language with full tool support and a methodology for mining such models.

Michael presented the talk on subject "Fast and Accurate Declare Mining".

Abstract:
Declarative models make it possible to describe changing systems or systems with a lot of freedom by imposing global constraints instead of explicitly defining the flow of control.  State-of-the-art process mining of declarative models either limits the types of constraints that can be mined, impose assumptions that are not always realistic, spend prohibitively long time on the mining process, or several of these.  For example, contemporary Declare mining cannot cope with a process with two sub-processes, one rare and expensive, and one common and cheap.  Contemporary approaches would be unable to discover anything meaningful for the arguably more important expensive sub-process.

In this presentation, I present a highly efficient means to mine Declare models.  Our approach does away with all previous assumptions and limitations, and instead checks these.  This means we can potentially construct much better models as we have much more information available.  This allows us to also provide sensible data in the case with the two sub-processes.  We use a number of existing and new techniques to achieve this, including symmetry reduction, parallelization, superscalar mining, and prefix sharing.

Significantly improved speed (from days or weeks to seconds or minutes) also makes it possible to use Declare mining as a sub-procedure for more general mining algorithms.  For example, we can implement the Alpha algorithm from scratch in minutes (and extend it to cope with error for free), or we can discover block-structured models with ease.

Everything presented is implemented in ProM and available now.

Attendees are expected to be familiar with process modeling and basic computer science like automata.  Concepts like declarative models with be explained in the talk.

Slides: see at author's personal page