=Paper=
{{Paper
|id=Vol-2973/paper_276
|storemode=property
|title=The Recomminder: A decision support tool for Predictive Business Process Monitoring
|pdfUrl=https://ceur-ws.org/Vol-2973/paper_276.pdf
|volume=Vol-2973
|authors=Christoph Drodt,Sven Weinzierl,Martin Matzner,Patrick Delfmann
|dblpUrl=https://dblp.org/rec/conf/bpm/DrodtWMD21
}}
==The Recomminder: A decision support tool for Predictive Business Process Monitoring==
https://ceur-ws.org/Vol-2973/paper_276.pdf
The Recomminder: A Decision Support Tool for
Predictive Business Process Monitoring
Christoph Drodt1 , Sven Weinzierl2 , Martin Matzner2 and Patrick Delfmann1
1
Institute for IS Research, Universität Koblenz-Landau, Koblenz, Germany
2
Institute of Information Systems, Friedrich-Alexander-Universität Erlangen-Nürnberg, Nuremberg, Germany
Abstract
Predictive business process monitoring (PBPM) provides a set of techniques to optimize the performance
of operational business processes. Most recent PBPM techniques learn predictive models from historical
event log data using machine learning algorithms (ML). However, there is no silver bullet approach
for different event logs, and their performance depends on the characteristics of the underlying event
logs. This paper demonstrates the decision support tool Recomminder. The main idea of our tool is
to recommend an appropriate pre-processing procedure, an ML algorithm, and the hyper-parameter
configuration for a new event log based on its characteristics. While our tool can support researchers
to better understand the relation between event log characteristics and ML-driven PBPM techniques, it
supports practitioners in developing effective PBPM techniques.
Keywords
Predictive Business Process Monitoring, Machine Learning, Business Process Management, Process Min-
ing, Decision Support
1. Introduction
Over the last years, business process management (BPM) researchers have developed a plethora
of predictive business process monitoring (PBPM) techniques. PBPM techniques aim to predict
aspects such as next activities, process outcomes, or next timestamps in running business
processes. Based on these predictions, process stakeholders can proactively intervene in run-
ning business processes. In doing that, process stakeholders can improve the performance of
operational business processes by mitigating risks or avoiding failures before these occur, or
exploiting potentials in time [1].
Most recent PBPM techniques generate predictions through predictive models learned from
historical event log data using machine learning (ML) algorithms [2]. Driven by the goal to
achieve accurate predictions, many techniques with different data pre-processing procedures,
ML algorithms, and hyper-parameter configurations have been proposed [3, 4].
However, PBPM research has shown that there is no silver bullet approach for event logs
Proceedings of the Demonstration & Resources Track, Best BPM Dissertation Award, and Doctoral Consortium at BPM
2021 co-located with the 19th International Conference on Business Process Management, BPM 2021, Rome, Italy,
September 6-10, 2021
" drodt@uni-koblenz.de (C. Drodt); sven.weinzierl@fau.de (S. Weinzierl)
~ http://fg-bks.uni-koblenz.de/ (C. Drodt)
0000-0002-4682-8036 (C. Drodt); 0000-0003-2268-7352 (S. Weinzierl); 0000-0001-5244-3928 (M. Matzner)
© 2021 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org)
Event logs
Training
Log
Trainer Evaluator
Analyzer
Decision
Recommending
New Event log Support
Log Recom-
Analyzer mender
Figure 1: Four Components of the Recomminder
with different characteristics, even if predictive models are learned using deep learning (DL)[5].
Instead, most approaches are typically optimized based on a specific set or type of event logs
but often show a poor generalization ability if applied to other event logs. In general, event logs
differ regarding their characteristics, e.g., the number of activities, number of trace variants, or
number of categorical data attributes. So far, little is known about the relation between event
log characteristics and ML-driven approaches in BPM research. As a logical consequence of
this missing understanding, it is challenging for practitioners to adopt an approach designed
and implemented by research because their available event logs may comprise characteristics
to that such an approach is generally not geared.
To overcome this challenge, we demonstrate the decision support tool Recomminder in this
paper for the task of predicting next activities. The tool is mainly structured into the four
components LogAnalyzer, Trainer, Evaluator, and Recommender (see Fig. 1). For a given set
of event logs, the LogAnalyzer component first determines event log characteristics. Second,
the Trainer component learns predictive models for different constellations of pre-processing
procedures, ML algorithms, and hyper-parameter configurations. Third, the Evaluator tests the
predictive models and calculates ML metrics to assess the models’ predictive quality. Forth, the
Recomminder component selects per event log the predictive model with the highest predictive
quality and creates a meta predictive model in form of a decision tree (DT) that learns the
mapping from the event log characteristics to the best performing (non-meta) predictive models.
Finally, the meta predictive model can be applied to a new event log to determine an appropriate
pre-processing procedure, ML algorithm, and hyper-parameter configuration based on its
characteristics.
2. Backend
In this section, we explain the structure of the Recomminder and elaborate its technical back-
ground. The presented tool is written in Python 3.8.2. and is designed as a Python package,
split into two main modules: backend (this section) and frontend (Section 3). The backend is
organized into four components, namely Log Analyzer, Trainer, Evaluator and Recommender,
which will be described in more detail in this section. Advanced users can access these com-
ponents and their modules by simply importing them and, thus, accessing their methods in
their code. Additionally, we designed two phases that represent the fundamental workflows
of Recomminder: the offline phase (Training/Feeder) and the online phase (Recommending)
(see Fig. 1 and Section 1). On the coding perspective, those phases can be executed by calling
Recomminder.train(), and Recomminder.recommend(). To make the features of the Recomminder
also available for non-developers, we designed a web-based frontend that can be started via the
shell or command line. For more details about the frontend, please see Section 3.
Another important aspect is to allow other developers to extend this artifact, especially the
Trainer component, by implementing more classifiers. We developed a superclass to support
developers in implementing further classifiers. More details on executing, extending, and
installing the Recomminder can be found in the repository (see Section 4).
2.1. Log Analyzer
In terms of the feature extraction, we consider three different types of event log characteris-
tics: event-log-based (e.g., concerning events, activities, and traces), process-model-based (e.g.,
concerning loops, noise, variants, and gateways), and process-context based features (e.g., con-
cerning categorical attributes and numerical attributes). To extract some of these features, a
Process Discovery algorithm has to produce a process model. In this tool, we used PM4PY1 ’s
data structure for event logs and its Heuristic Miner [6] implementation to create a Petri net
of a event log file. The Heuristic Miner is a common discovery algorithm in process mining
(PM) that can, e.g., handle loops and noise in event log data [7, 8]. After feature extraction, the
results are stored in an SQLite database for later use.
2.2. Trainer
To evaluate the best classifier for a given event log file, a predictive model must be trained
for each classifier. In the Recomminder tool, we implemented two classifiers using existing
ML frameworks: Random Forest2 [9] and LSTM3 [10]. The presented tool also provides some
methods for pre-processing the event log. This includes event encoding methods (ordinal
and one-hot encoding for categorical attributes and min-max normalization for continuous
attributes), as well as sequence encoding techniques (window-based and index-based prefix
generation). In addition, we implemented a test-train-split method to generate test and train
data sets. Before executing the training, we evaluated the best fitting hyper-parameters using
the Optuna4 framework with 20 optimization rounds and a subset of the training set (90%
training and 10% test split). Finally, the results are stored in the database. After each classifier
has been trained, the model is transferred to the Evaluator.
1
https://pm4py.fit.fraunhofer.de
2
https://scikit-learn.org/
3
https://www.tensorflow.org
4
https://optuna.org
2.3. Evaluator
This component evaluates different common metrics (accuracy, precision, recall and f1 score)
and stores them in the database. Once all Trainer processes have finished, the best classifier
for an event log file is extracted and used as target values. In addition, the Evaluator gathers
corresponding extracted features of the log which are used as input samples. Those data is the
training set for the DT. The training of the DT is performed by the Evaluator and the resulting
model represents the meta predictive model for the Recommender. Finally, we retrieve feature
importance values from the trained meta model, that is the outcome of this component, and
store a visualisation of the evaluation.
2.4. Recommender
As stated in the previous section, the Recommender components rely on a DT which is trained
with the results from the Log Analyzer and Trainer. By calling the Recomminder.recommend()
method, the Recommender uses the Log Analyzer to extract the feature of a given event log file
and feeds them to the trained DT model. Following, the best matching classifier is returned.
3. Frontend
In addition to the developer access level, we also provide a rich, web-based frontend to allow non-
developers to use the Recomminder. To implement the frontend, this tool uses TurboGears25 for
content delivery, Bootstrap6 to style the web pages, and jQuery7 for asynchronous functions.
The navigation is split into the two workflows described previously and is located at the top of
the page. Stepping into one phase, the tool provides a sub-navigation that leads the user through
the necessary steps. In both phases, users can upload event log files via an asynchronous upload
script, so that multiple event logs can be provided easily. The frontend directly interacts with the
backend and starts both functions (train end recommend) as a new thread. Using the threading
extension of python, the frontend can proceed without waiting for the new process to finish.
Further on, the frontend shows a live view on the Recomminder’s log file, so users can always
follow current process steps and retrieve the process. Finally, the frontend presents the results
and in case of the training phase, it also includes three figures, including a representation of
the DT, a plot of the metrics, and a bar chart stating the feature importance. Every progress is
stored in the session, generated by TurboGears2, which allows users to return to the frontend
after some time and pick the process up where they left it.
4. Conclusion
This paper demonstrated the decision support tool Recominder, consisting of four components
that can support research and practice. With our tool, researchers can better understand the
relationship between event log characteristics and ML-driven PBPM techniques. On the other
5
https://turbogears.org
6
https://getbootstrap.com
7
https://jquery.com
hand, our tool can support practitioners in the development of effective PBPM techniques. In
future research, we plan to extend the tool by:
• further prediction tasks such as the next timestamp prediction,
• further granularities of prediction tasks such as prediction per prefix size or prediction
per decision point,
• further ML algorithms to learn predictive models,
• further sequence and event encoding techniques,
• further intrinsic explainable ML algorithms such as linear regression to learn the mapping
between event log characteristics and the best-performing predictive models, and
• ensemble learning techniques such as stacking or boosting to improve the prediction
accuracy by combining several predictive models.
A demonstration video of the Recomminder can be found at https://youtu.be/37ikmt9g818.
The source code is of the decision tool is available under the Lesser GNU Public License (LGPL)
at https://gitlab.uni-koblenz.de/fg-bks/predictive-recommining.
References
[1] A. E. Marquez-Chamorro, M. Resinas, A. Ruiz-Cortes, Predictive Monitoring of Business
Processes: A Survey, IEEE Transactions on Services Computing 11 (2018) 962–977.
[2] C. Di Francescomarino, C. Ghidini, F. M. Maggi, F. Milani, Predictive Process Monitoring
Methods: Which One Suits Me Best?, in: Business Process Management, volume 11080,
Springer International Publishing, 2018, pp. 462–479.
[3] K. Heinrich, P. Zschech, C. Janiesch, M. Bonin, Process Data Properties Matter: Introducing
Gated Convolutional Neural Networks (GCNN) and Key-Value-Predict Attention Networks
(KVP) for Next Event Prediction with Deep Learning, Decision Support Systems 143 (2021)
113494.
[4] S. Weinzierl, S. Zilker, J. Brunk, K. Revoredo, A. Nguyen, M. Matzner, J. Becker, B. Es-
kofier, An Empirical Comparison of Deep-Neural-Network Architectures for Next Activity
Prediction Using Context-Enriched Process Event Logs, arXiv (2020).
[5] W. Kratsch, J. Manderscheid, M. Röglinger, J. Seyfried, Machine Learning in Business
Process Monitoring: A Comparison of Deep Learning and Classical Approaches Used for
Outcome Prediction, Business & Information Systems Engineering 63 (2021) 261–276.
[6] A. Weijters, W. M. P. van der Aalst, A. K. Alves de Medeiros, Process Mining with the
HeuristicsMiner Algorithm, Citeseer, 2006.
[7] A. P. Kurniati, G. Kusuma, G. Wisudiawan, Implementing heuristic miner for different
types of event logs, International Journal of Applied Engineering Research 11 (2016)
5523–5529.
[8] P. Weber, B. Bordbar, P. Tino, A principled approach to mining from noisy logs using
Heuristics Miner, in: 2013 IEEE Symposium on Computational Intelligence and Data
Mining (CIDM), IEEE, 2013, pp. 119–126.
[9] L. Breiman, Random Forests, Machine Learning 45 (2001) 5–32.
[10] S. Hochreiter, J. Schmidhuber, Long Short Term Memory. Neural Computation, Neural
Computation 9 (1997) 1735–1780.