Process mining has proven its usefulness in a wide range of practical applications, however, it generally requires event logs to be certain to guarantee trustworthy results. Typical approaches to uncertainty in event logs, like frequency-based filtering or other heuristics, restrict insight into processes because they possibly discard relevant information. Several alternative approaches to uncertainty in process mining have been proposed to provide more detailed insights, but these approaches each only address a single type of uncertainty. A great number of domains could benefit from process mining techniques that can handle the simultaneous occurrence of multiple types of uncertainty, e.g., probabilistic processes where event logs are extracted from unstructured data. The proposed PhD project aims to develop a holistic approach to uncertainty in process mining, adding a comprehensive perspective of uncertainty to the insights generated by process mining analyses. To achieve this, methods concerned with diferent types of uncertainty in process mining, namely data, correlation, and process uncertainty, will be investigated, and then combined into a harmonized framework, providing a foundation for improved decision-making.
Process mining has proven its usefulness in a wide range of practical applications in order to
uncover bottlenecks and ineficiencies in processes or to identify tasks for automation [
Generally, three diferent types of uncertainty exist for processes discovered from unstructured data, namely data uncertainty, correlation uncertainty, and process uncertainty. Data uncertainty refers to the degree of noise in the data like inaccurate, imprecise, untrustworthy, and unknown data. Correlation uncertainty refers to the likelihood of event-activity mappings since there is often more than one possible solution. Then, process mining can also be afected by the uncertainty inherent in the analyzed process (i.e., probabilistic dependencies and contextual influences afected by randomness).
The reduction and quantification of uncertainty might improve process discovery results, improve conformance checking and predictive monitoring and even elevate process discovery techniques on unstructured data. The purpose of this PhD project is to develop a holistic framework to address uncertainty in process mining, especially process mining applied to unstructured data. This includes quantifying the impact of uncertainty, uncovering the sources of uncertainty in event log extraction, quantifying the random factors of uncertainty when mapping events onto activities, and providing a user-friendly communication of the uncertaintyaware process mining methods.
In most process mining approaches, uncertainty is treated as an issue of event log quality, and addressed with frequency-based filtering or other heuristics. While this is a practical approach to reduce noise (i.e., erroneous recordings), it may also suppress outliers (i.e., correctly recorded, but unexpected behavior), which are highly relevant to analyze process deviations [5].
Recently, some approaches have been suggested that integrate uncertainty into process mining, instead of disregarding uncertain event data. Pegoraro [6] proposed a framework that adds a perspective on data uncertainty to process mining. An event log is annotated with (meta-)information related to the uncertainty of the events contained in an event log, which is used as additional input to uncertainty-aware algorithms for process discovery and conformance checking. Also, process models are annotated to represent the uncertainty of the event log they were discovered from. Qafari et al. [7] proposed an approach to identify the causes of process performance and compliance problems from event logs. Structural causal models are used to discover causal relations between distinct features (e.g., event or case attributes, the occurrence of an activity) and problematic process outcomes. Leemans et al. [8] developed a method to identify long-term dependencies between control-flow decisions in a process. Control-flow decisions are identified from a process model and event log of this process, then probabilistic causalities between control-flow decisions at the decision points are discovered, and finally, the size of each causal efect is estimated. Alman et al. [ 9] proposed a framework to extend declarative process mining methods with a process uncertainty perspective.
To sum up, no approach exists to quantify diferent types of uncertainty in process mining. Mostly, existing approaches are limited to either data or process uncertainty and focus on settings where structured event logs are available. The integration of multiple types of uncertainty into process mining in a combined approach is still an open challenge. Additionally, current event log extraction techniques are unable to provide explicit uncertainty information, leaving a blind spot with respect to correlation uncertainty. Thus, handling uncertainty is particularly challenging for process mining on unstructured data, where event logs need to be extracted ifrst.
A large body of research is available on the quantification of uncertainty in domains other than process mining (e.g., deep learning [10], mechanical engineering [11], or climate modeling [11]), which can be built on to provide uncertainty quantification techniques for process mining, especially in order to address data and correlation uncertainty. For instance, Zhang et al. [12] provide a general framework guiding the application of existing uncertainty quantification methods. Abdar et al. [10] discuss applications of uncertainty quantification in deep learning. Similarly, to address process uncertainty, the PhD project can rely on a large body of causal inference techniques [13] to quantify probabilistic dependencies in processes.
III
II
I
IV
V
The goal of the PhD project is to address the various challenges outlined above. Particularly, a structured approach making uncertainty explicit in the process discovery phase and quantifying the degree of multiple types of uncertainty will be developed. The focus of this approach lies on the analysis of unstructured data since the level of uncertainty in the data is very high and all three types of uncertainty are significant. Fig. 1 shows the general structure of approaches to process mining on unstructured data (e.g., [14]), which are used to divide the PhD project into five phases. Phases (I)-(III) serve to enrich event logs with information on uncertainty and its impact on the trustworthiness of the insights, and phases (IV) and (V) are concerned with developing uncertainty-aware process mining methods and communicating the impact of uncertainty related to the gained insights. To support the generalizability of our approach, large and heterogeneous evaluation datasets with known ground-truth processes and characteristics (e.g., the amount of uncertainty) will be created by generating synthetic event data [15].
The first step addresses data uncertainty in event logs (I). For this, a taxonomy of uncertainty in event logs will be developed and used to generate synthetic event logs of certain processes with varying levels of data uncertainty. Then, uncertainty quantification [ 12, 10] methods will be adapted to assess the impact of data uncertainty related to the quality of insights gained from the process mining results. Next, the scope is widened to include correlation uncertainty by integrating uncertainty awareness into event log extraction (II). The applicability of the methods developed in (I) will be extended to unstructured data sources by developing extraction techniques that produce explicit uncertainty information. The quantification of data uncertainty can be extended to event log extraction techniques to enable the automatic identification of sources of data uncertainty. In the third step, process uncertainty is addressed (III). Process uncertainty can be isolated by generating high-quality event logs of uncertain, probabilistic processes. To provide a solution, (1) existing approaches for discovering (probabilistic) causalities in process mining will be reviewed, (2) causal inference techniques [13] will be adapted to address the gaps identified in this review, and (3) means to enrich event logs with quantitative information of the discovered causalities will be developed. The goal of the fourth step is to integrate the diferent views on uncertainty explicitly into common process mining tasks (IV). To do this, uncertainty-aware process mining techniques need to be developed, which can explicitly encode uncertainty to improve the quality compared to non-uncertainty-aware techniques and improve the quantification of uncertainty through measures. Finally, explainability will be addressed to improve the communication to non-experts between the technical design and the process mining results (V). The explainability of uncertainty provides a basis to gain additional insights for informed process management decisions. For this, diferent means to communicate process mining results (e.g., process models, reports) need to be ofered for the uncertainty-enriched outputs developed in the previous phases.
In this PhD proposal, the challenges related to diferent types of uncertainty in process mining were described. The goal of the proposed PhD project is to address these challenges by suggesting a holistic framework to manage uncertainty in process mining. In order to achieve this, methods to enrich event logs with information on data, correlation, and process uncertainty will be developed, and this information will then be made explicit in the analysis results by integrating uncertainty into process mining methods and the communication of process mining results. The PhD project contributes to trustworthy process mining, laying the foundation for improved process mining-based decision-making.
I would like to thank my thesis advisor, Agnes Koschmider, for her invaluable guidance and support in writing this proposal, and Milda Aleknonytė-Resch for the helpful discussions about the research design and presentation of the proposal. This project has received funding from the German Federal Ministry for Economic Afairs and Climate Action under the Marispace-X project grant no. 68GX21002E.