Dirk Fahland (TU/e)

New research challenges for statistics, probability theory, and optimization in process mining

The field of Process Mining aims at getting insights into processes and organizations by building understandable models from event logs. The field originated from formal language theory, specifically Petri nets, an input event logs as a finite (weighted) sample of an infinite formal language, and analysis as optimization problems over graphs that can be constructed from the input. Problems are primarily solved by heuristic search algorithms.

In my research, I have developed different ways of modeling the input event log: as weighted (directed-acyclic) graphs of various forms to encode how time progresses over various entities/objects/actors recorded in the data. These graphs give rise to interesting problem formulations that ask for new solution methods drawing from statistics, probability theory, and various forms of optimization in networked data.

In this talk, I will present 3 exemplary problems:

1) I present a probabilistic formulation of the “process discovery problem” for constructing the most likely model of the data generating process that produced the sample and a first sequential learning approach to solve it that – I believe – can still be improved significantly.

2) I present a model of event data that explicitly encodes time and movement of objects in time over “work stations”. This model describes how items move through queueing network. Visualizing the model reveals complex patterns of overtaking, grouping, and where density and “speed”

of items vary or deviate from their surroundings; these deviations in one part of the data are correlated (and can be sometimes explained as causally related) to deviations in other parts of the data, but we currently have no reliable methods for detecting patterns and their correlations/cause-effect relations.

3) I present a network-based model of items interacting with each other over time. I discuss the properties of the model and various new kinds of questions that arise in this data that have close links to (social) network analysis but have never been formulated or approached in this way in the process mining field.

Through these 3 examples, I would like to invite the audience to study new problem areas, possibly identify novel kinds of problems, and hopefully develop a joint interest in solving (some of) them.

—–

This talk will be held in a hybrid fashion; you are welcome to attend the talk on-campus in MF15, or via Zoom. The Zoom link is attached to the bottom of this invitation.

For upcoming events of the SPOR seminar, as well as a history of previous talks, see: https://www.eurandom.tue.nl/eindhoven-spor-seminar/. Recordings of earlier talks are available upon request.