II. RESEARCH OBJECTIVES AND MOTIVATION
Process complexity can be defined based on four main
characteristics [
7
]: structure, interconnection, dynamics and
uncertainty. To a certain extent, these properties are
necessary for business processes in order to operate in
volatile and complex business environments. But the
complexity of a business processes alone does not explain
how well a process is prepared to face such operational
challenges. Hence, instead of only looking at process
complexity in general, one has to look at how complexity
contributes to process diversity. In system theory, diversity is
considered as a strategy to make a system more resilient
against uncertainty, to support innovation and to produce
better outcomes over the long run [
3
]. An important
difference to the general concept of process complexity is,
that diversity also considers the degree of dissimilarity among
a set of objects and not simply, whether there exist
differences among objects. In this PhD project we want to
explore this viewpoint, by analyzing process diversity from
multiple perspectives – in particular control flow, resources
and data – and by using diversity as guidance for business
process improvement. Hence, the objectives are threefold:
1) Define process diversity measures, which reflect the
dissimilarity of process traces, based on multiple
perspectives.
2) Use the defined diversity measures in combination with
performance measures to discover reference models for
process improvemenr.
3) Empirically investigate the impact of business process
diversity on process performance.
While process diversity in terms of variability of
execution traces, has not been specifically investigate in BPM
research so far, a lot of research has focused on the
complexity of business process models [
8
]. Several measures
have been developed, which consider the number of process
model components and their interconnectivity, to assess the
overall complexity of the model. However most of these
measures only focus on the control-flow perspective and none
of the measures actually considers process diversity. By
adapting some of these measures, such that they can be
applied for the assessment of event logs, it will be possible to
also assess other process perspectives and to reflect diversity
to a certain extent (objective 1). We will also investigate, how
to extend existing measures, such that they can be applied to
object-centric event logs [
9
], since they provide very detailed
information regarding the interconnection of process objects.
To further extend the existing measures and to find new ones,
we will also consider diversity theory [
4, 5
]. Diversity theory
has been applied in different fields such as biology and
economics, with the general purpose to assess the
dissimilarity among a set of objects, e.g. populations or
product portfolios. The existing measures in these fields, can
provide different perspectives on diversity, depending on the
underlying assumptions. For example, whether dissimilarity
is understood as a cardinal or ordinal concept, can
significantly change the outcome of the assessment. Hence, a
thorough analysis and evaluation of the different approaches
is necessary, to develop suitable diversity measures for
business processes.
The next step of this PhD project (objective 2) is to use
the identified measures in combination with fitness and
additional performance measures as quality metrics for
process discovery. In this way it will be possible to discover
reference models, for business process standardization,
which provide a certain degree of diversity and therefore
decreases the risk of negative side-effects and failure. Similar
to [
10
] we will use genetic process mining in order to
combine different quality metrics for process discovery.
However, we will use a different evolutionary approach,
called quality-diversity algorithm [
6
]. Due to an efficient
exploration of the given search space, the approach decreases
the risk of convergence towards local optima and it allows to
discover multiple process models, with an equally high
diversity level. These models can then be analyzed and
selected for process standardization based on preferences
regarding control-flow, resources and data. A further
important benefit of the quality-diversity algorithm is that this
approach is able to develop different scenarios, how a process
might evolve in the future and to show whether a process is
able to adjust to changing conditions, in terms of resilience.
In order to provide further guidance on process
standardization, we will empirically investigate the impact of
diversity on business process performance and evolution
(objective 3). Based on the developed diversity measures we
want to show how diversity evolves over time and how this
impacts long-term process performance. Eventually this leads
to a theoretical model, which is able to explain, to which
extend and under which circumstances process diversity is
able to improve performance.
III. RESEARCH METHODOLOGY
The first step of the PhD project is to identify existing
process complexity measures that can be applied to process
models and event logs. As a starting point we use the
literature review on process model complexity provided by
Polančič and Cegnar [
8
]. The existing measures will be
analyzed and adapted, such that they reflect the multiple
perspectives of process diversity. We will also look at other
related research fields, such as software engineering and
project management to identify potentially new measures. In
addition, we will adapt measures from diversity theory [
4, 5
],
in order to capture the dissimilarity between traces and events
in an event log. The evaluation of the identified measures will
be based on defined mathematical axioms, which ensure
consistent results across different data. In addition, the
measures will be compared based on standard as well as
object-centric event logs. Some of the event logs will be
artificially generated, such that different parameters can be
adjusted (e.g., number and interdependence of resources and
activities) in order to see how well these changes are reflected
by the measures. In order to see whether, the measures can
actually be used for the analysis of complexity in reality, we
will also use real event log data from BPI challenges.
In order to use the defined measures as guidance for
business process standardization, we apply them as quality
measures for business process discovery. In particular we will
use the quality-diversity algorithm [
6
] to discover multiple
reference models with a high diversity. We will compare the
results with other process discovery algorithms, which are also
able to consider multiple quality dimensions (e.g., [
10
]). The
comparison will be based on the level of diversity, which the
discovered models exhibit and how well the algorithms
perform in terms of computation. For this purpose, we will use
publicly available event logs from past BPI challenges to
create a benchmark.
These event logs will also be used for the empirical
analyses part of this PhD project. We will first use cluster
algorithms to identify different process variants based on
changes over time and measure the evolution of diversity
accordingly. We will additionally use statistical inference to
analyse the impact of diversity on business process
performance.
IV. RELATED RESEARCH
[
11
] is the only study we could identify, which also relates
event log complexity to process discovery. The study
proposes to use graph entropy as a measure for complexity.
This measure only refers to the control-flow complexity and
does not consider process diversity. However, the approach
can potentially be extended to incorporate this perspective.
Based on this extension it would be interesting to analyse how
different process discovery algorithms are able to reflect
process diversity.
There also exists some research regarding business
process standardization based on event logs. One of the
earliest attempts is described in [
12
]. However, the main goal
of this approach is to discover a reference process model, with
a minimal distance to the existing process variants. If a process
exhibits very high diversity in terms of control-flow, resources
and data, the approach most likely does not reflect this in the
reference model. In contrast, our approach allows to discover
multiple reference models, which also consider diversity and
performance.