Process analysis often follows an exploratory approach: while scrutinizing the event data captured from a process, analysts continuously build hypotheses and subsequently seek to falsify or validate them based on the data. This potentially involves comparing different process mining algorithms or testing different parameters. However, the exploration of the event data does not just start with the analysis, but rather when the data is prepared, as the data extraction, transformation, and loading (ETL) may also be subject to frequent change. Thus, decoupling the analysis from the data preparation possibly hides the effect of the choices made during the ETL steps on the analysis outcome. That does not only make it difficult to relate the results of the analysis to the original data, but also to judge the reliability of the results at large. Database technology appears to have the means in store to address that challenge, as they allow to integrate the ETL process and analysis using unified data models and query languages. Instead of extracting the event data to a log, one may keep the data close to the source, thus allowing to trace the analysis results to the source data. Ultimately, this affords the opportunity to reason on the propagation or interplay of changes in the pre-processing phase with respect to the analysis. That way, process analysts are supported in tracking and understanding the impact of decisions made during data preparation and analysis, which enables them to justify those decisions. Furthermore, adopting standard data models and query languages as the basis for this integration enables us to leverage the capacity of database systems and the research on them spanning decades to support the process analysis. Accordingly, the complex of problems that is going to be addressed in the PhD thesis can be summarized as follows:

We aim at achieving traceability in process analysis, i.e., the ability to trace analysis results to the input data. To accomplish this, we will integrate the models used for data pre-processing and data analysis, and operationalize them using existing data management capacities.

ICPM 2024 Doctoral Consortium, October 14-18, 2024, Kongens Lyngby, Denmark maike.basmer@hu-berlin.de (M. Basmer) 0000-0003-2124-0276 (M. Basmer)

Within the relational realm, intermediate in-database representations and a native in-database operator have been developed to accelerate process mining tasks [1,2]. Furthermore, concepts from data warehousing were adopted to facilitate multidimensional analysis [3,4]. Schönig et al. [5,6] implemented declarative process discovery on relational databases. Riva et al. [7] considered different schemata that have been proposed to represent event logs in the past and examined the effect of the schema choice on the performance of declarative process mining. Besides that, modelling event data as labeled property graphs [8] was proposed to enable graph-based understanding of multi-dimensional data and to accommodate different analyses [9].

A process analysis pipeline may encompass different pre-processing steps like integrating, transforming, reducing, abstracting, filtering, or enriching the event log before the analysis [10], with abstraction currently being the focus for our setting. Different approaches to event abstraction exist, mainly lifting low-level events to activities according to the domain [11,12]. Other types of high-level events may also be discovered to enhance the analysis of processes [9,13].

Regarding traceability, there have been several proposals in the past. Probabilistic event abstraction allows to keep track of alternative abstractions by capturing uncertainty when producing high-level events [14]. For process mining on IoT data, Bertrand et al. [15] propose a schema for an event log that caters to traceability concerns as well as different needs in granularity. Klinkmüller et al. [16] examine the sensitivity of discovery results w.r.t. pipeline operations and parameters to debug process discovery pipelines, encompassing the discovery procedure itself along with pre-processing steps like abstraction or filtering. Data and provenance views were proposed to support explorative process mining by tracking steps, goals, and intermediate results throughout the analysis process [17]. Beyond process mining, further inspiration may be drawn from research on provenance [18], explanations [19], debugging of pipelines [20], or probabilistic databases [21].

In the course of the PhD project, several facets may be investigated, for example:

• Which data schema or data model should be used depending on the use case or the characteristics of the data? • Can we exploit properties of the data to support the process analysis?

We will focus on two specific use cases described below to grasp and better understand these questions and the arising challenges.

To target the traceability of abstractions and their effect on a given analysis, the concept of Event Knowledge Graphs (EKGs) [8] implemented in graph databases [22] may come in handy, as they integrate low-level events with high-level abstractions and enable graph-based querying. This capacity may be extended to record event abstractions, such that the effect of abstractions during exploratory process analysis can be tracked. To that end, we conceive the following framework: In a forward-manner, the abstractions represented as queries in a given data preparation pipeline are treated as first-class citizens of an EKG by recording them along with their relations to lower-level events. Considering an alternative abstraction in the pipeline, the intermediary results of that alternative pipeline are computed and recorded correspondingly. Differences in the analysis may be explained by the difference set of nodes or edges between both possible "worlds" -either by their mere (non-)existence in one set or the other or by the context they define (i.e., the features that distinguish those nodes or edges). We plan to apply this idea to a pipeline for task analysis [23], as it involves several steps of abstraction. Interaction mining [24] may also lend itself to evaluating this idea.

Similarly, the rich feature set of relational database systems may be employed to host process mining tasks. We plan to focus on declarative process mining [25], especially in view of multiple dimensions [26,27], as data-aware conditions relate to selection and navigating relations correspond to joins in the relational model. Implementing conformance checking or process discovery for multi-dimensional declarative process specifications encompasses finding an adequate representation of the event data, encoding the task as a set of queries, and ideally leveraging database technology like materialized views [28] to track and reuse intermediary results. Another aspect that could be exploited in case of declarative process specifications is their apparent similarity to data dependencies in relational databases. In that case, techniques from the domain of data profiling may be used as a basis for, e.g., the discovery of declarative constraints [29]. Beyond that, it might be interesting to investigate which intermediary data representations like indices [30] or materialized views [28] or other developments from database systems research like row pattern recognition [31] may be useful to realize process mining tasks in-database.

Developments aiming at enhancing the efficiency of process analysis tasks may be evaluated empirically in a set of experiments on data sets that are established within the process mining community. In addition to that, synthetic data may serve to investigate the influence of specific data properties on the interventions that are going to be devised during the PhD project. When it comes to evaluating the traceability, one can either head into the direction of showing that the developed approach fulfills certain properties or measure the capacity of the proposed approach to trace deviations in the analysis due to abstractions. For example, it might be sensible to measure how compact these insights can be represented if we assume a correlation between the compactness of the representation and understandability.

The proposed thesis sets out to integrate pre-processing of the event data with process analysis by means of database technologies to achieve traceability. We outlined ideas how to approach this problem set, e.g., through the lens of event knowledge graphs (in terms of database technology used) or declarative process mining (in terms of process analysis).