The Process Analytics Workbench (PAW) is a comprehensive tool to streamline Predictive Process Monitoring (PPM) activities. It integrates functionalities like synthetic event log generation, model training, and prediction evaluation within a micro-service architecture, enabling agile and eficient experimentation.
Value
Predictive Process Monitoring (PPM) is a subcategory of process mining that is concerned
with the prediction of the future state of process instances, which are currently being executed.
Common predictions are, for example, the next events or the duration till the process is
completed. PPM tools usually rely on Machine Learning (ML) algorithms trained on event log data
[
Our goal is to provide a tool that supports all major aspects of PPM. We therefore propose the Process Analytics Workbench (PAW), which supports all the main PPM functionalities out of the box and can easily be extended to meet the specific application requirement. Through components including the model training, predictions and synthetic log generation, experiments can be quickly set up and conducted with little efort.
PAW consists of a collection of micro-services, each fulfilling a single purpose. This design approach is central to the tool, providing many advantages for end users. Users can determine the modules which are relevant to their application and run the corresponding micro-services. This approach optimizes time and resource utilization while maintaining minimal complexity in the PPM environment. Figure 1 shows an overview of the currently available modules.
Process
Model Generator
Simulator
Connector
Trainer
Predictor
The currently available micro-service modules can be broadly categorized into three major groups, namely the process model generation and simulation, the data connector, and the training and prediction module.
The process model generation and simulation module allows users to quickly and easily create batches of diferent process models and simulate the respective event logs. It provides vast configurability to support a large number of experimental purposes.
We developed the process model generator based on an algorithm proposed by F. Bezerra et
al.[
The process model simulation is realized by integrating the agent-based workflow simulator
proposed by Amann and Delfmann [
To enhance reproducibility, random seeds used for stochastic functions in the process model generation and event log simulation modules can be set with presets, enabling researchers to replicate experiments without additional efort.
The data connector provides all process data to the trainer and predictor. It takes raw process data from a source and transforms it into a standardized format that the trainer and predictor can use. The current data connector is developed to extract event log data from the Camunda 8 workflow management system. This makes the integration into an environment running on Camunda 8 efortless. The main idea behind the connector is to provide a single piece of software for the provision of process data.For other use cases, a modified version of the connector could be developed to connect to a diferent workflow management system.
The training and prediction modules constitute the core components of PAW, facilitating the
training of predictive models and the prediction of attributes for ongoing process instances. Data
preprocessing described in [
To demonstrate PAW’s potential application scenarios, we conducted a preliminary experiment evaluating the correlation between the control flow complexity (CFC) of a process model and the performance of a predictive model trained using event logs simulated based on this process model.
Cardoso [
4. Resource complexity: resources required to execute the activities and the process. First, we randomly generated 25 process models, each containing a fixed number of 10 tasks. No data dependencies or information flows were assigned to these process models. The predictive task was set as the next activity prediction, ensuring that activity, data flow, and resource complexities were excluded, leaving only CFC to influence the ground truth for prediction. The 25 process models were categorized into five groups based on their control flow complexity, with values of 1, 3, 5, 7, and 9. Event logs with 200 cases each were simulated from the process models with same simulation parameters.
In the subsequent phase, XGBoost models with identical hyperparameters were trained and evaluated with the simulated event logs, employing a split ratio of 64:16:20 for training, validation, and testing.
We conducted experiments on five groups, each comprising five process models, and consolidated the results into mean and standard deviation values, reflecting the distribution of the results in each group.
Model CFC
Based on the provided results, a general observation can be made that as the process model CFC increases, the predictive model’s performance tends to decline in terms of precision, recall, and F1 score. However, an exception to this trend is observed between the CFC values of 5 and 7. This deviation suggests that the relationship between control-flow complexity and predictive performance may not be strictly linear, and other factors might influence the results.
During the literature review, we could not identify publicly available BPMN generators with
comprehensive documentation besides PLG2[
We also identified other popular and publicly available BPMN model simulators, namely BIMP2 and Apromore3, and compared them to our tool. Table 2 provides an overview of the main diferences. The major distinction between our BPMN simulator and others lies in the basic approach of the simulation engine. While the established simulators rely on the propagation of simulated tokens through the process model, our approach simulates the whole environment in which the process runs. This provides many more configuration options and allows for a high complexity in the simulation. For example, while the established simulators select a path after an XOR gateway by defining fixed probabilities for each path, our approach handles the selection based on the values of variables in the simulated environment. This allows for very high flexibility and simulates dynamic scenarios, including concept drift. However, because of the added complexity, our simulation approach takes several minutes, while the established simulators usually finish in several seconds.
Simulation Approach Simulation Complexity
Simulation Time
BIMP Token-Based
Low Fast
Apromore Token-Based
Low Fast
PAW Agent-Based
High Slow
No comparable applications within the Business Process Monitoring (BPM) domain have been identified that simplify and automate the machine learning pipeline for PPM. As demonstrated in the case study, our tool enables rapid experiment execution, produces statistically robust results, and efectively addresses randomness issues in experimental design.
PAW provides a comprehensive solution for PPM activities. It supports many diferent application scenarios, from industry utilization to research. The micro-service architecture provides plenty of opportunities for customization and future expansions. Our case study showed how PAW can facilitate and support experiments in the PPM domain.
However, this work only represents the initial development iteration of the tool and does not constitute a final and complete state of the software. Currently, its complex installation process is not friendly for users without solid computer and programming knowledge. Also, the current user interface does not allow for the control of all aspects of the tool. We plan to address these shortcomings with the upcoming releases. We also plan to expand the existing platform further to include additional features in the form of new micro-service modules, such as hyperparameter optimization and data clustering. 2https://bimp.cs.ut.ee/ 3https://apromore.com/ This research is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 456415646, 432399058.