=Paper=
{{Paper
|id=Vol-2420/papeDT10
|storemode=property
|title=Online Comparison of Streaming Process Discovery Algorithms
|pdfUrl=https://ceur-ws.org/Vol-2420/papeDT10.pdf
|volume=Vol-2420
|authors=Kavya Baskar,Marwan Hassani
|dblpUrl=https://dblp.org/rec/conf/bpm/BaskarH19
}}
==Online Comparison of Streaming Process Discovery Algorithms==
https://ceur-ws.org/Vol-2420/papeDT10.pdf
Online Comparison of Streaming Process
Discovery Algorithms
Kavya Baskar1,2 and Marwan Hassani1,3
1
Department of Mathematics and Computer Science
Eindhoven University of Technology, Eindhoven, The Netherlands
2
k.baskar@student.tue.nl
3
m.hassani@tue.nl
Abstract. In the active field of process mining, several techniques have
been proposed in various areas like process discovery and conformance
checking. The integration of data stream mining techniques in process
mining has gained popularity in recent years. The ProM framework that
enables process mining with streaming data has been advanced to sup-
port event streams in the recent past. In this paper we present a new ex-
tension that is built upon existing work related to obtaining process mod-
els from data streams within ProM. The extension enables researchers
to visually compare the results of two different process discovery algo-
rithms for a single incoming stream of events with different algorithms to
deal with the data streams such as Lossy Counting with Budget, Sliding
Window and Exponential Decay.
Keywords: Process mining · Event streams · ProM · Visualization
1 Introduction
We assume the reader to be acquainted with the field of process mining and
refer to [5] for a detailed understanding of the field. The concept of process
mining in the streaming domain has become a subject of interest among re-
searchers over the recent years. To enable the application of process mining on
event streams, several techniques have been proposed with regard to data stor-
age, process discovery algorithms, conformance checking and event stream based
process enhancement. One such development focuses on abstract representation
approximations using algorithms which are designed for frequent item mining
on data streams [8]. A prototypical implementation that corresponds to this
development is provided in the process mining tool-kit ProM4 .
The previous architecture is a generalization and standardization of existing
event stream-based process discovery algorithms and defines a computational
mechanism applicable to a large class of process discovery algorithms. The gen-
eralization of the architecture allows for the inclusion of event stream-based
process discovery algorithms within the framework, in the future.
4
http://www.promtools.org/
2 K. Baskar and M. Hassani
This demo is based on this existing implementation of process mining with
streaming data, in the ProM tool. Henceforth, the remainder of this demo will
focus towards the ProM framework [6]. Previous implementations on the ProM
tool mainly allow the user to obtain the resulting process model for only one of
process discovery algorithms [7] [2]. In our proposal we provide researchers and
practitioners with the ability to compare visually the resulting Petri nets from
two specific process discovery algorithms. The idea of visually comparing the
output of two different streaming algorithms with different settings or the same
algorithm with different parameter settings at each time stamp is inspired by the
stream clustering tab of MOA (Massive Online Analysis) framework [1] which
is an open source framework for data stream mining with concept drift. The
aim of this demo is to enhance the evaluation of streaming process discovery
algorithms within the ProM framework by enabling an online analysis of the
resulting models. Hence, the design decision was made to be able to compare
only two algorithms simultaneously.
2 Architecture
ProM is an open source framework for a wide variety of process mining algo-
rithms and techniques in the form of plug-ins. It is implemented in Java and is
therefore platform independent.
A plug-in in ProM is an algorithm implementation that is of significance
which agrees with the framework 5 . For this demo, we have created a new plu-
gin based on five existing packages within the ProM code base, i.e. Stream,
EventStream, StreamAbstractRepresentation, StreamAlphaMiner and StreamIn-
ductiveMiner. The new plugin is implemented in the package KavyaBaskar 6 that
has dependencies on the aforementioned packages which have been modified to
suit the new architecture.
The feature of the proposed architecture in this demo is the visualization of
the resulting process models (Petri nets) from two different event stream-based
process discovery algorithms. In [7] a standardized approach that extends ProM
framework enabling the handling of streaming data is presented. In [8] an ar-
chitecture is proposed which allows for the adoption of several process discovery
techniques in the event stream context, making it very generic. The core of the
implementation in this demo builds upon the architecture presented in [8]. The
two specific process discovery algorithms chosen for this implementation are the
Alpha Miner and the Inductive Miner. Alpha Miner was chosen since it is the
very first miner to have been created for process discovery, and the Inductive
Miner because it guarantees a sound workflow net and previous researches in-
dicate it to be one of the best process discovery algorithms currently available.
A new plug-in called CompareStreamInductiveMinerStreamAlphaMinerAPNX-
SEventReaderImpl has been created in the ProM framework and the respective
5
http://www.processmining.org
6
https://www.dropbox.com/s/xko9gbntjgclrhy/eclipse-workspace.rar?dl=0
Online Comparison of Streaming Process Discovery Algorithms 3
algorithms have been incorporated within this plug-in. Other combinations are
currently under development.
The new implementation provides the additional possibility of choosing the
internal data structures for each miner- Backward Exponential Decay, Forward
Exponential Decay, Lossy Counting with Budget and Sliding Window apart from
the existing choice between Frequent, Lossy Counting and Space Saving.
The implementation of the Visualization object has been extended to split
the display into two panels with process models resulting from the two different
algorithms displayed on either panel. As present in earlier implementation, the
Slider component at the bottom of the panel and the start/pause/stop buttons
can be used to view the older model and analyze the evolution of the models.
Two separate Slider components have been implemented in the new plug-in to
enable the viewing of older models in any order and to be able to compare the
models generated by the mining algorithms at different points of time. Moreover,
the Update Result button can be used to generate model(s) at any random point
of time as long as the stream of events are incoming.
3 Case study
As an explanatory case we have used the BPI Challenge 2017 data set7 for the
stream generation. A detailed demonstration is provided as a screen cast 8 .
The log is imported and a stream is generated using the Generate Event
Stream plug-in, setting a default emission speed of 10 data-packets/second. An
example of the resulting stream is depicted in Figure 1a. A tutorial for using the
(a) Stream Generator object visualization (b) Applying the process discovery algo-
of active stream. rithms on the live event stream.
Fig. 1: Generating event streams and applying the algorithms
tool is available 9 . While the stream is being generated, the CompareStreamInductiveMiner-
7
https://doi.org/10.4121/uuid:5f3067df-f10b-45da-b98b-86ae4c7a310b
8
https://www.youtube.com/watch?v=2eMZe6NWhW0&feature=youtu.be
9
https://www.dropbox.com/s/yh2rzymgy3higqn/Tutorial_Online_Comparison_
of_Streaming_Process_Discovery_Algorithms_%28BPM_2019_Demo%29.pdf?dl=0
4 K. Baskar and M. Hassani
StreamAlphaMinerAPNXSEventReaderImpl plug-in is selected as shown in Fig-
ure 1b, with Event Stream (XSEvent) as input which is created by the Generate
Event Stream plug-in. In this plug-in, the Case and Activity Identifier(s) can
be specified. From the drop down list of data structures, the Forward Expo-
nential Decay has been chosen for the Inductive Miner with Renewal Rate=1,
Threshold=0.01 and Decay Rate=0.01. Lossy Counting with Budget for the Al-
pha Miner with Budget size=1000 has been chosen. The result is depicted in
Figure.2.
Fig. 2: Resulting Petri nets sample of two event stream-based process discovery
algorithms.
4 Future Work
We plan to include conformance checking metrics such as fitness and precision
comparison for both the models, visualized using a time series graph, in order to
compare the performance of both algorithms. The implementation is in progress,
as illustrated in Figure 3. Near future goal is to incorporate the prospect of se-
lecting any two of the existing algorithms from the event stream-based process
discovery domain(e.g. StrProM [3]), with the possibility of including future ones
too (see [4] for the potential of sequential pattern mining approaches for stream-
ing process discovery). Possibility of implementing this architecture in other
streaming frameworks such as Storm, Spark, Flink, Kafka and Samza can be
explored.
5 Conclusion
The newly presented simultaneous visual analysis of process models for the same
event stream empowers researchers, developers and business users to experiment
Online Comparison of Streaming Process Discovery Algorithms 5
Fig. 3: Implementation of conformance checking metrics visualization.
with the concept of process discovery with streaming data within the process
mining domain. The extension also allows the user to analyze the internal data
structure used for handling the data stream within the ProM framework.
The lessons learned during the development of the presented implementation
can be used to tackle the hurdles in building a more generic and standardized ar-
chitecture which will hopefully enable the complete implementation to compare
all process models resulting from algorithms of choice.
References
1. Albert Bifet, Geoff Holmes, Richard Kirkby, and Bernhard Pfahringer. MOA: Mas-
sive online analysis. Journal of ML Research, 11(May):1601–1604, 2010.
2. Andrea Burattin. Process mining for stream data sources. In Process Mining Tech-
niques in Business Environments, pages 177–204. Springer, 2015.
3. Marwan Hassani, Sergio Siccha, Florian Richter, and Thomas Seidl. Efficient process
discovery from event streams using sequential pattern mining. In IEEE Symposium
on Computational Intelligence and Data Mining (CIDM), pages 1366 – 1373, 2015.
4. Marwan Hassani, Sebastiaan J. van Zelst, and Wil M. P. van der Aalst. On the
application of sequential pattern mining primitives to process discovery: Overview,
outlook and opportunity identification. WIREs Data Mining and Knowledge Dis-
covery, e1315, 2019.
5. Wil Van Der Aalst. Process mining: discovery, conformance and enhancement of
business processes, volume 2. Springer, 2011.
6. Wil MP van der Aalst, Boudewijn F van Dongen, Christian W Günther, RS Mans,
AK Alves De Medeiros, Anne Rozinat, Vladimir Rubin, Minseok Song, HMW Ver-
beek, and AJMM Weijters. ProM 4.0: comprehensive support for real process anal-
ysis. In Intl. Conf. on Application and Theory of Petri Nets, pages 484–494, 2007.
7. Sebastiaan J van Zelst, Andrea Burattin, Boudewijn F van Dongen, and
HMW (Eric) Verbeek. Data streams in ProM 6: A single-node architecture. In
BPM (Demos), page 81. Citeseer, 2014.
8. Sebastiaan J van Zelst, Boudewijn F van Dongen, and Wil MP van der Aalst.
Event stream-based process discovery using abstract representations. Knowledge
and Information Systems, 54(2):407–435, 2018.