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 diferent 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 efective PBPM techniques.
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
running 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 [
Most recent PBPM techniques generate predictions through predictive models learned from
historical event log data using machine learning (ML) algorithms [
However, PBPM research has shown that there is no silver bullet approach for event logs g n ii n a r T g n i d n e m m o c e R
with diferent characteristics, even if predictive models are learned using deep learning (DL)[
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 diferent 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.
In this section, we explain the structure of the Recomminder and elaborate its technical background. 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 components 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 ofline 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).
In terms of the feature extraction, we consider three diferent types of event log
characteristics: 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.,
concerning 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 [
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[
1https://pm4py.fit.fraunhofer.de 2https://scikit-learn.org/ 3https://www.tensorflow.org 4https://optuna.org
This component evaluates diferent 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.
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.
In addition to the developer access level, we also provide a rich, web-based frontend to allow nondevelopers 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.
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 5https://turbogears.org 6https://getbootstrap.com 7https://jquery.com hand, our tool can support practitioners in the development of efective 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.