Digitization as the megatrend of the 21st century has the potential to in uence many aspects of human life [2]. Based on recent technological advances and the ensuing growth in maturity, Arti cial Intelligence (AI) can be seen as the main driving factor behind this development. It is expected to revolutionize the ways in which humans work, learn, communicate, consume, and live. From a business point of view, digitization and AI o er great potential, while simultaneously posing a signi cant risk [4]. In the last decade, many companies were outperformed by their more digitized competitors, causing severe losses or even bankruptcy. With its potential for automation and innovation, AI is expected to continue or intensify this development. On the other hand, the market for AI software applications is expected to grow tremendously in the coming years [1]. Business leaders are often aware of these developments, but unsure how to leverage the potential of AI for their own business processes. There is, hence, an ongoing need for AI research, both to develop new methods and technologies and to transfer them into entrepreneurial practice. Modern and digitized business processes are centered around data, which not only triggers its execution, but also in uences the decisions that lead towards the overall process goal [13, 4]. Given the advanced digitization of business processes and the widespread availability of data, business process management (BPM) is a well-suited eld for AI application. Most current research focuses on AI for process execution. However, BPM contains other tasks, which could also benet from AI either automating laborious tasks and freeing more human capacity or allowing new and previously impossible insights into the process. The available process data is a good starting point for BPM researchers to develop new AI methods that also support process development, modeling, implementation, monitoring, and optimization. This is the vision of this thesis, as formulated in the guiding question: How can AI technologies be applied for BPM? Concretely, the thesis investigates the application of AI technologies in three exemplary BPM subtopics at di erent maturity stages regarding both research and practical adoption: Reference Model Mining (RMM), Predictive Process Monitoring (PPM), and Process Discovery (PD). The di erent starting points,
in addition to the di erent goals and data availability, provide researchers with di erent challenges that require di erent solutions and the use of di erent technologies. This shows both the variety and the spectrum of opportunities of applying AI in BPM. 2
For the topic of RMM, the main research question focuses on the lack of
maturity and practical adoption of currently existing techniques, calling for the
development of new methods to overcome this obstacle (RQ 1: How can current
RMM methods be enhanced to foster their practical adoption?). The question
is divided into three di erent subquestions, regarding the state-of-the art in
RMM and its advancement by means of AI. For the rst subquestion (RQ 1.a:
What are current challenges in inductive reference model development?), we
analyzed the current state-of-the-art in inductive reference modeling and reference
model mining in order to identify critical challenges and apparent research gaps
[
Some challenges can be overcome by widening the scope of RMM techniques
to derive reference models not only from other individual process models, but
also from instance-level data that record the behavior of information systems.
Compared to type-level data, instance-level data is larger and more widely
available, which are important requisites for applying AI methods. Therefore, for the
second subquestion (RQ 1.b: How can a reference model be mined from event log
data?), we developed two new approaches for mining reference models from event
logs [
Another emerging shortcoming is the lack of methodical support for nding
the right RMM approach (RQ 1.c: How can a suitable RMM method be selected
for a given application case?). Di erent approaches will produce di erent
reference models for the same input data, without providing any guidance on where
and how this reference model should be used. To overcome this challenge in the
practical adoption of RMM, we developed the concept of \Situational
Reference Model Mining", i.e., the idea that the intended reference model purpose
determines the requirements to a reference model and, therefore, the approach
that is best suited for mining it [
To address the application of AI in BPM for RMM, we focused on methods that combine automated approaches with human intelligence to achieve better results with fewer resources. The tools analyze and structure the available input data according to rational criteria. Their data-centric view assists the human reference model developer, who is able to also take soft factors into account. In this regard, our contributions do not directly target AI in a narrow sense. Instead, they foster the collaboration of humans and systems, where each party brings its individual assets towards the solution of the problem. 3
The guiding research question for PPM is centered around the application of
state-of-the-art deep learning technology for predicting the next events in a
process sequence (RQ 2: How can deep learning techniques be used to develop new
methods for PPM?). It is separated into two subquestions, regarding the
development of a new method for next-event prediction and its practical application
and enhancement. In order to address the rst subquestion (RQ 2.a: How can
deep learning be used to predict the next events in a process sequence?), we
presented a novel approach to predicting the next process event using deep learning
[
Our contribution to PPM is the design of a new AI system in form of a trained neural network. It is novel in the sense that it does not receive any information about explicit process structure. Instead, it independently develops an implicit understanding of the structure and is able to reason about future process behavior. Users do not have to provide any information beyond a process log, but they also cannot access the implicit representation of the process behavior except when making predictions.
The main research question for PD addresses the validity of process discovery evaluations, separated into two subquestions (RQ 3: Which in uences may threaten the validity of process discovery evaluations?). This question di ers from the other two, because it is a knowledge question instead of a design question. It is motivated by existing artifacts (process discovery methods and quality metrics) and contributes to the investigation of a problem context, such that eventually, new design questions can be asked, and new artifacts can be created.
The rst subquestion (RQ 3.a: How is the quality of process discovery
evaluations in uenced by unobserved process behavior?) is concerned with the
inuence potential unobserved behavior in an event log might have on the quality
assessment of the discovered model. We conducted an empirical study, which
examines how the epistemological problem of induction (generalizing singular
observations) in uences established notions of measuring the quality of the
discovered models [
The second subquestion (RQ 3.b: Which mistakes can be made during a
process discovery evaluation that will compromise its validity?) is motivated
by the results of the rst one. If established measures for process discovery
quality do not factor in the in uence of unobserved behavior, then there might
also be other threats to their validity. After identifying a list of 20 potential
threats (hyperbolically named \process mining crimes"), we perform a literature
review to determine their prevalence [
Our research on PD did not focus on the development of new AI systems, but pursued the notion of rationality in process mining. What constitutes a \good" process discovery result and how can we measure it? Our goal was to establish a common notion of process discovery quality, which can be used by humans and AI systems alike. When comparing process discovery approaches with potentially di ering rationality functions, this quality notion can be used as a measure for meta-rationality, allowing researchers to develop new approaches and evaluate the existing ones. Hence, our contributions to PD facilitate the development of new AI-based methods for process discovery, while simultaneously enabling human process miners to compare and improve the quality of their own approaches. 5
As there are in nite possibilities to use AI in BPM, this summarized thesis can only provide some suggestions. Illustrating the demonstrative nature of our results, we exemplarily examined three BPM subtopics and how they could bene t from AI technology. They provide readers with a rst idea of how diverse AI research might be even in a comparably narrow eld like BPM. For each subtopic, we either introduce new methods that advance its state-of-the-art or gain new knowledge, which may enable the development of such methods in the future.
AI in general and machine learning in particular have multiple limitations and their application to BPM is no exception. Machine learning approaches, particularly deep learning, need large amounts of training data to produce viable results. This data has to be collected, recorded, stored, and checked, which can become a practical challenge when working with real-life process systems. In addition, AI systems are unable to handle unknown situations or take \soft factors" into account. In general, AI is well suited for solving a particular class of problems. Whether or not a particular problem in BPM falls into this class, remains to be decided on a case by case basis. Research on AI for BPM can take many di erent forms and purposes, all of which intend to contribute to the advancements of business processes and the companies that run them.