Operational Process Mining
In this section, we provide a literature survey of operational process mining.2 Specifically,
we go over three streams of work, namely simulation mining, supervised process mining,
and multi-perspective conformance checking. The former two lines of work focus on
performance-oriented operational questions, while the latter focuses on the conformance
between model and log.
Simulation Mining. Early works on performance-oriented process mining were based
on the enhancement of control-flow models with further data-based information such as
decisions, resources and time [
6
]. The work of Ma˘rus¸ter and van Beest [
7
] extends this
approach from modeling and analysis of an existing process to its improvement based
on the discovered model. Due to the richness of these discovered models they can only
be analyzed via simulation. Thus, we refer to these methods as simulation mining or
model-based process mining, interchangeably.
The main benefit of these model-based techniques is their ability to solve a large
set of performance problems. For example, one can use simulation to predict waiting
times and propose resource staffing. However, current simulation mining techniques
overlook queueing effects by simulating process cases independently of each other and
drawing waiting times (and other measures that stem from case dependencies) from
fitted distributions. Furthermore, model-based mining techniques tend to overfit reality
as it is observed in the event log and therefore it was shown that simpler performance
models often provide accurate results and are more efficient [
8
]. The expressiveness
problem is partially addressed in [
4
], where authors mine non-Markovian stochastic
Petri nets [
9
], which are simple formalisms that account for control-flow and exponential
durations. In [
8
], the best of both worlds is combined by a method for simulation mining
that captures queueing effects and simplifies the resulting model. This simplification
results both in improved efficiency (run-time complexity of the solution) and accuracy.3
Supervised Process Mining. Another prevalent stream of work in process mining is the
encoding of event logs into well-established supervised learning methods. Examples
for such methods can be found in [
3,10,11,12
]. In these works, both the behavior and
the operational aspects of the process are encoded: the former representing the
controlflow perspective and the latter capturing activity durations and resources. While these
approaches are often accurate in solving a given mining problem with respect to a set of
performance measures observed in the event log (e.g. total time in process), it has two
major drawbacks.
First, one cannot use these techniques for quantifying measures that are not directly
observable in the log. For example, information that allows one to calculate resource
utilization might be missing from the event log, and thus supervised approaches cannot
be applied. Second, exploring process improvement directions and sensitivity of process
parameters (e.g. durations and arrival rates) is impossible due to the lack of data that
2 Solving operational problems based on data is not limited to process mining. For references to
works outside the process mining discipline, see [
4
] (for operations research), and [
5
] (for data
mining).
3 The work is a result of the current research. It is not included in this thesis.
describes the effect of changing these parameters. In [
5
] and [
13
], we propose
supervised process mining solutions that bring together the benefits of model-based mining
(or simulation mining) and supervised learning. Specifically, these solutions comprise
machine learning methods (e.g. decision trees) and queueing models. This combination
enables quantifying measures that were not observed in the event log due to the existence
of the model. For measures that appear in the event log we get accurate solutions by
exploiting the strengths of supervised learning [
14
]. Furthermore, we are able to have
what-if reasoning based on the queueing model, without the need to acquire further data
from the changed process.
Multi-Perspective Conformance Checking. In this thesis, we call operational process
mining to works on multi-perspective conformance analysis that verify alignment of
process models and event logs with respect to time and resources [
15,16
]. The approach in
the literature is to compare deterministic data values according to a given model to those
available in the log. For example, activity durations are considered to be known values
(according to the model) and the alignment procedure checks whether or not event log
durations are equal to model durations. However, the reality in processes is often random
(random durations, random arrival process, etc.). Therefore, a methodology for
comparing distributions (that come from the model) and deterministic values (coming from
the log) is required. We show such an approach in our work on conformance checking
from the queueing perspective [
17
]. Further, our work proposes a methodology that
considers queueing interactions that were neglected in existing methods, and recommends
improvement based on the result of the conformance checking procedure [
17
].
3
Thesis Contributions
The main overall contribution of this thesis is the development of queue mining, which
is a set of operational process mining solutions based on the discovery of queueing
models from event data. These solutions explicitly capture delays due to case-resource
interactions. Below, we detail the specific contributions of this research and outline their
impact in two research fields, namely process mining and queueing theory (or, more
broadly, operations research).
3.1
Contributions to Process Mining
The thesis makes the following contributions to process mining:
– The work integrates data-driven queueing modeling and analysis into operational
process mining [
13,17,5,23,8,22
], namely queue mining.
– Our work links operational conformance analysis to performance improvement in
business processes [
17
]. Specifically, we develop a methodology for the detection of
problematic areas in a process (e.g., due to synchronization or resource delays) and
propose proper adjustments to queueing policies (selection of the next case to enter
service) in these problematic parts.
– We make advances in the area of multi-dimensional conformance analysis, by
adding the queueing perspective, structuring the approach, and demonstrating how
conformance checking can lead to performance improvement [
17
].
– We propose hybrid methods that combine supervised learning and model-based
process mining techniques to improve the accuracy of answering operational
questions [
13,5
]. This combination improves accuracy with respect to applying the two
approaches separately.
Three of the works mentioned in the list above, [
13,17,5
] are part of the thesis. Other
works were published in papers that were not included in the thesis. The proposed
methods were validated based on three real-world datasets coming from service domains.
In [
22,13
] we evaluate queue mining against a bank’s call center; in [
17,23,8
] we mine
the dataset of the Dana-Farber Cancer Institute (a large outpatient cancer hospital in the
United States); and in [
5
] we evaluate our techniques on GPS bus data that comes from
the city of Dublin.
3.2
Contributions to Queueing Theory
Queueing theory has been mainly a theoretical research field that aims at modeling and
analyzing (typically) stochastic and dynamic service or manufacturing systems [
24,25
].
Some works in queueing theory provide numerical experiments that justify the
practicality of theoretical results. The underlying data is typically simulated based on systems
that adhere to various assumptions that fit the models. Only a small sample of works
in queueing theory include real-world data validation (see [
26,27,28
] and references
within). In this work, we validate multiple results from queueing theory, thus showing
their practical applicability to real-world data. Further, we provide a platform for
testing new and (other) existing results in queueing theory based on event data, and the
queueing model is well-defined). Specifically, all our techniques were implemented as
prototype systems that are available as open-source. For example, the methods that we
use for delay prediction [
13
] can be extended, replaced, and tested against, when new
prediction methods become available. In other words, this research creates a playground
for queueing theorists to empirically evaluate their developed theories and results.