In the active eld 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 support event streams in the recent past. In this paper we present a new extension that is built upon existing work related to obtaining process models from data streams within ProM. The extension enables researchers to visually compare the results of two di erent process discovery algorithms for a single incoming stream of events with di erent algorithms to deal with the data streams such as Lossy Counting with Budget, Sliding Window and Exponential Decay.
We assume the reader to be acquainted with the eld of process mining and
refer to [
The previous architecture is a generalization and standardization of existing event stream-based process discovery algorithms and de nes a computational mechanism applicable to a large class of process discovery algorithms. The generalization of the architecture allows for the inclusion of event stream-based process discovery algorithms within the framework, in the future.
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 [
ProM is an open source framework for a wide variety of process mining algorithms 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 signi cance which agrees with the framework 5. For this demo, we have created a new plugin based on ve existing packages within the ProM code base, i.e. Stream, EventStream, StreamAbstractRepresentation, StreamAlphaMiner and StreamInductiveMiner. The new plugin is implemented in the package KavyaBaskar 6that has dependencies on the aforementioned packages which have been modi ed 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 di erent event stream-based
process discovery algorithms. In [
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 di erent 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 di erent 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
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
algoof active stream. rithms on the live event stream.
tool is available 9. While the stream is being generated, the
CompareStreamInductiveMiner7 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
StreamAlphaMinerAPNXSEventReaderImpl plug-in is selected as shown in
Figure 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 Identi er(s) can
be speci ed. From the drop down list of data structures, the Forward
Exponential 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
Alpha Miner with Budget size=1000 has been chosen. The result is depicted in
Figure.2.
We plan to include conformance checking metrics such as tness 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
selecting any two of the existing algorithms from the event stream-based process
discovery domain(e.g. StrProM [
The newly presented simultaneous visual analysis of process models for the same event stream empowers researchers, developers and business users to experiment 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 architecture which will hopefully enable the complete implementation to compare all process models resulting from algorithms of choice.