Process mining builds a bridge between traditional process modelling and data mining. Process discovery algorithms aim at constructing process models automatically from event logs. Existing process discovery algorithms perform well with simple event logs, but their performance can be afected when the input event logs contain complex behaviours, and the discovered process models may not represent the real behaviours of the business processes. In this research, we plan to develop methods to automatically detect diferent complex behaviours in event logs. The detection of complex behaviours can be based on the extensions of existing process discovery algorithms or stand-alone complex behaviours detection tools. The paper presents our research questions, methodologies, current research progress and potential challenges.
Process mining is a relatively new subject which builds a bridge between traditional process
modeling and data mining [
Many process discovery algorithms have been proposed, and some can return process
models with high fitness and precision values [
To discover comprehensive process models, many event log filtering tools have been proposed
[
As a goal of process mining is to facilitate business process improvement taking the advantage of event logs, it is important for us to get deep understanding about business processes. Filtering out and ignoring complex behaviours in event logs is inappropriate.
To this end, the proposed PhD project focuses on the detection and analysis of diferent complex behaviours in event logs. The research questions are presented below: • Can we extend existing process modeling languages to present diferent complex behaviours? • Can we extend existing process discovery algorithms to discover diferent complex behaviours? • Can we develop stand-alone tools to discover diferent complex behaviours directly from event logs?
On the one hand, existing process discovery algorithms can be extended to handle complex behaviours. By extending existing process discovery algorithms, their advantages can be inherited. On the other hand, stand-alone tools to discover diferent complex process behaviours can be developed. These tools can be applied before performing data pre-processing and process discovery algorithms to help us understand the complex behaviours ignored by the discovered model.
The rest of this paper is structured as follows: Section 2 is a literature review of related work. Section 3 presents our research approach. Section 4 presents our current research progress. Section 5 presents potential challenges of our research project and concludes the paper.
The original alpha algorithm [
The inductive miner [
Many methods have been developed in order to cluster traces in event logs [
Besides trace clustering algorithms, some methods also focus on the discovery of hierarchical process models to handle complex event logs with many activities and sub-processes (eg., [19, 20]). A high-level process model can be used to represent the relationship between diferent sub-processes while a low-level model can be used to represent the process model of each single sub-process.
To deal with changes within business processes, various methods have been proposed to
deal with concept drifts in event logs [
Some research also focuses on detecting other complex behaviours. For example, Lu et al.
[24] propose a method to handle duplicate activities in event logs. More specifically, the same
activity may execute in diferent stages of a process. A more comprehensive process model
can be discovered if we split such an activity into multiple activities. Dees et al. [
In a nutshell, stand-alone tools are independent from process discovery algorithms. They can discover more insights about the process from event logs which usually cannot be discovered directly by process discovery algorithms. These methods can be used together with process discovery algorithms to enhance our understanding of business processes described in event logs.
Our research approach is adapted from the Design Science Research (DSR) [26] methodology. Our research contains the following stages: • Identifying Problems: At this stage, literature review is conducted. The literature review focuses on research related to detecting process behaviours from event logs (eg., process discovery algorithms, concept drift detection algorithms, trace clustering algorithms, etc.). Based on the literature review, gaps can be identified among existing literature (eg., existing concept drift detection algorithms cannot distinguish process drifts from noises). • Designing and Developing Solutions: Once a gap has been identified among existing literature, a solution can then be proposed to solve the problem. Then a software can be built to implement the proposed solution. Existing open-source software packages can be modified for faster development process. • Demonstration and Evaluation: When a potential solution is implemented, evaluation should be conducted to show the capabilities of the proposed solution. The evaluation usually contains two parts: firstly, the method is evaluated empirically using a big number of synthetic datasets. The performance of the proposed method can also be compared with existing methods. Secondly, the method is evaluated using real-life datasets. • Communication: If a valid solution is built, the work can be presented to the research community through academic conferences, workshops or journals.
In this section, we briefly introduce some of our work to discover complex process behaviours in event logs. Please refer to the corresponding references for the details of our proposed methods and evaluation results. It has to be noted that all work presented in this section has been evaluated using publicly available datasets. The implementations of these methods are also publicly available.
In [17], we propose a novel method to extend the inductive miner to discover switch behaviours. Assume there is an exclusive decision choice in a process model, and the decision point is split into multiple branches. In real-life event logs, it is possible to switch between diferent branches after the decision has been made. However, due to the limitation of process trees, the original inductive miner is unable to discover such behaviours. A ”follower model” with very low precision can be returned when such behaviours exist in event logs.
To solve the problem, we firstly extend the process tree notation to allow the presentation of switch behaviours. The new notation is called ”switch process tree”. Each switch process tree can be translated into an equivalent workflow net. In addition, with some constraints of switch process trees, their corresponding workflow nets can be guaranteed to be sound. Fig. 1 shows an example process tree and its corresponding workflow net.
The extension of process trees allows us to extend the inductive miner to discover switch behaviours. As shown in Table 1, when switch behaviours exist in event logs, our proposed method can significantly improve the precision of the discovered process models. The discovered process models by the extended inductive miner can accurately present the process behaviours in event logs.
Finally, the method is also evaluated using a real-life dataset (”BPIC13-incident” event log from the ”4TU Center for Research Data”). As shown in Table 2, a more accurate and simpler process model is discovered by our proposed method.
In [21], we propose a method to accurately detect time points of process drifts in event logs. A process drift point is defined as the time point when there is a statistically significant diference among the observed process behaviours before and after the change point. Most existing process drift detection algorithms assume the input event logs to be clean and cannot diferentiate process drifts from noises. Our proposed method can not only accurately detect process drift points, but can also be robust to noises. Evaluation results show that our proposed method can consistently perform better than existing methods. Fig. 2 shows the evaluation results comparing to the baseline [22] when 20% of noises are inserted into the event logs.
In [28], we propose a method to detect branching frequency changes in process models. Branching frequency changes refer to changes in frequencies between diferent options when there is an exclusive choice. Our method is evaluated using a publicly available event log (”Italian Help Desk” event log from the ”4TU Center for Research Data”). The process model is presented in Fig. 3 and the results are presented in Fig. 4. Two frequency changes are detected in the event log, and the frequencies between choosing activity ”Wait”, ”Require upgrade” and ”Resolve ticket” after ”Take in charge ticket” change significantly after each drift point.
In this paper, we explain the need for developing new techniques to discover complex behaviours in event logs for better understanding of business processes. On the one hand, existing process discovery algorithms can be extended to handle complex behaviours. On the other hand, stand-alone tools can also be constructed for the detection of specific types of behaviours.
Although we have provided some solutions for the research problem, there are still some challenges. Firstly, there is the lack of methods to validate the results. For example, in [17], although a process model with higher precision value can be obtained from the publicly available event log, it may still be insuficient to conclude that switch behaviours actually exist in the process. We plan to evaluate this work empirically to evaluate its capabilities in the future. Secondly, some methods heavily rely on user-defined parameters. For example, In [ 21], a window size is required from users. The window size defines the number of events in each sample when performing statistical tests. Diferent process drift points can be reported with diferent window sizes. In [28], the detection of drift points relies on an external change detection algorithm. The detection results are dependent on the parameters of the external algorithm.
Finally, we also plan to develop methods to discover other complex behaviours such as the automatic detection of cancellation behaviours and non-free-choice behaviours in event logs.
This research project is supervised by Associate Professor Simon K. Poon from the School of Computer Science, the University of Sydney. Springer International Publishing, Cham, 2017, pp. 93–113. [17] Y. Lu, Q. Chen, S. Poon, A novel approach to discover switch behaviours in process mining, in: S. Leemans, H. Leopold (Eds.), Process Mining Workshops, Springer International Publishing, Cham, 2021, pp. 57–68. [18] M. Leemans, W. M. Van Der Aalst, M. G. Van Den Brand, Recursion aware modeling and discovery for hierarchical software event log analysis, in: 2018 IEEE 25th international conference on software analysis, evolution and reengineering (SANER), IEEE, 2018, pp. 185–196. [19] X. Lu, A. Gal, H. A. Reijers, Discovering hierarchical processes using flexible activity trees for event abstraction, in: 2020 2nd International Conference on Process Mining (ICPM), 2020, pp. 145–152. doi:1 0 . 1 1 0 9 / I C P M 4 9 6 8 1 . 2 0 2 0 . 0 0 0 3 0 . [20] S. J. Leemans, K. Goel, S. J. van Zelst, Using multi-level information in hierarchical process mining: Balancing behavioural quality and model complexity, in: 2020 2nd International Conference on Process Mining (ICPM), 2020, pp. 137–144. doi:1 0 . 1 1 0 9 / I C P M 4 9 6 8 1 . 2 0 2 0 . 0 0 0 2 9 . [21] Y. Lu, Q. Chen, S. Poon, A robust and accurate approach to detect process drifts from event streams, in: A. Polyvyanyy, M. T. Wynn, A. Van Looy, M. Reichert (Eds.), Business Process Management, Springer International Publishing, Cham, 2021, pp. 383–399. [22] A. Ostovar, A. Maaradji, M. La Rosa, A. H. M. ter Hofstede, B. F. V. van Dongen, Detecting drift from event streams of unpredictable business processes, in: I. Comyn-Wattiau, K. Tanaka, I.-Y. Song, S. Yamamoto, M. Saeki (Eds.), Conceptual Modeling, Springer International Publishing, Cham, 2016, pp. 330–346. [23] A. Yeshchenko, C. Di Ciccio, J. Mendling, A. Polyvyanyy, Visual drift detection for sequence data analysis of business processes, IEEE Transactions on Visualization and Computer Graphics (2021). [24] X. Lu, D. Fahland, F. J. H. M. van den Biggelaar, W. M. P. van der Aalst, Handling duplicated tasks in process discovery by refining event labels, in: M. La Rosa, P. Loos, O. Pastor (Eds.), Business Process Management, Springer International Publishing, Cham, 2016, pp. 90–107. [25] Y. Dubinsky, P. Sofer, Detecting the “split-cases” workaround in event logs, in: A. Augusto, A. Gill, S. Nurcan, I. Reinhartz-Berger, R. Schmidt, J. Zdravkovic (Eds.), Enterprise, BusinessProcess and Information Systems Modeling, Springer International Publishing, Cham, 2021, pp. 47–61. [26] A. R. Hevner, S. T. March, J. Park, S. Ram, Design science in information systems research,
MIS Quarterly 28 (2004) 75–105. [27] A. Augusto, R. Conforti, M. Dumas, M. La Rosa, A. Polyvyanyy, Split miner: automated discovery of accurate and simple business process models from event logs, Knowledge and Information Systems 59 (2019) 251–284. [28] Y. Lu, Q. Chen, S. Poon, Detecting and understanding branching frequency changes in process models, in: A. Augusto, A. Gill, S. Nurcan, I. Reinhartz-Berger, R. Schmidt, J. Zdravkovic (Eds.), Enterprise, Business-Process and Information Systems Modeling, Springer International Publishing, Cham, 2021, pp. 39–46.