Analyzing How Process Mining Reports Answer Time Performance Questions

Analyzing How Process Mining Reports Answer Time Performance Questions Carlos Capit´an-Agudo1(B),Mar´ıa Salas-Urbano1, Cristina Cabanillas1,2 , and Manuel Resinas1,2 1SCORE Lab, Universidad de Sevilla, Seville, Spain {ccagudo,msurbano,cristinacabanillas,resinas}@us.es 2I3US Institute, Universidad de Sevilla, Seville, Spain Abstract. The advances in process mining have provided process analysts with a plethora of different algorithms and techniques that can be used for different purposes. Previous research has studied the relationship between these techniques and business questions, but how process analysts use them to answer specific questions is not fully understood yet. We are interested in discovering how process analysts respond to specific business questions related to time performance. We have coded 110 answers to time performance questions in more than 60 process mining reports. As a result, we have identified 55 different operations with 137 variants used in them. We have analyzed the types of answers and their similarities, and examined how contextual information as well as existing process mining support may affect them. The results of the study provide an overview of the current state-of-practice to answer time performance questions and unveil opportunities to improve process mining tools and the way these questions are answered. Keywords: Process mining ·Time performance ·Qualitative analysis ·Quantitative analysis ·BPI Challenge ·Grounded Theory 1 Introduction Many process mining techniques and tools have been developed in the last years to assist the discovery, monitoring and improvement of business processes based on the event logs provided by the information systems that support them [1]. Each technique usually targets specific aspects of the processes, such as the existence and order of the process activities [15], the assignment and distribution of process participants [4], or the time performance of the process execution [18]. The importance acquired by process mining has also led to the development of methodologies, guidelines and case studies on how to perform process mining. These have mainly concentrated on understanding or guiding the use of process mining from a global perspective but they have not explored extensively how process mining analysts use these techniques to respond to specific business questions [14]. A better understanding of this matter can help to identify limitations in the approaches followed by the analysts to answer such questions. A good example of this are the limitations derived from the widespread use of the directly-follows graph as a way to analyze process execution [2]. It may also help to find common patterns that facilitate the building of reference guidelines to support them in their task. Finally, it may ease the identification of gaps between the features of process mining tools and what is actually done by analysts. In this paper, we conduct a systematic analysis of process mining reports aiming to discover what process mining operations (e.g. filtering, data manipulation, graphical representation) are used by process analysts to address a specific type of business questions, namely, time performance questions; and how these operations are related to each other as well as to the questions. Time performance questions refer to aspects like cycle time, waiting time, or bottlenecks. They constitute one of the most recurrent problems in process mining projects [10]. The data source of our analysis are the process mining reports submitted to the BPI Challenge (BPIC for short), an annual competition since 2011. Every year the challenge organizers publish a real-life event log provided by an organization together with specific business questions posed by the organization, so that the solutions provide them added value. Participants answer these questions or perform other complementary analyses and submit them in a report. Considering time performance questions, this includes 62 reports belonging to 4 different BPIC with a total of 110 answers. There are several reasons for choosing BPIC as the source of our study. First, they provide different perspectives on how to analyze the same data for the same question, which makes the answers more comparable. Second, they cover several analyst profiles: academics, students, and professionals. Third, the analyses in the reports are not undirected, but driven by specific questions posed by the organization, which is aligned with the way process mining is used in practice [10]. Fourth, as the reports are page limited, they collect the most important and conclusive information, avoiding distractions with irrelevant information. Finally, all reports analyzed are publicly available, which helps with the traceability and replicability of the results. We have analyzed these reports applying a four-step methodology following a mixed-methods research approach similar to the one in [14]. The results obtained include a catalogue of 55 operations and 137 variants that provides an overview of the current state-of-practice to answer time performance questions. The study also gives insights about the type of answers that can be found in the reports, how the context of the analysis (pursued goal, log and authors’ profile) affects the characteristics of the answers, and what is the observed impact of current state of the art on the answers. These results can be useful to further improve process mining tools and the way in which questions are addressed by analysts. The paper is structured as follows. Section 2outlines the literature related to this work. Section 3describes the methodology followed to conduct the analysis. Section 4provides details of the analysis and the findings. Section 5summarizes the conclusions drawn and directions for future work. 2 Related Work Methodologies and guidelines to do process mining have been developed over the last 10 years. Several methodologies define high-level stages, inputs, outputs and activities that should be performed in a process mining project. Examples are the Process Diagnostic Method [3], the L∗life-cycle model [1]andPM2[10], which identifies 6 stages in a process mining project: planning, extraction, data processing, mining & analysis, evaluation, and process improvement & support. Similarly, [12] provides guidelines to support organizations in systematically using process mining techniques aligned with Six Sigma. Due to their broad scope, these methodologies are intentionally open in terms of which techniques can be used to address specific questions. Instead, we are interested in understanding the details of how time performance-related questions are addressed in practice. The methodologies are useful though, to frame the context of the research presented in this paper: the mining & analysis and, partially, the data processing stage. Another workstream has focused on analyzing published process mining case studies to provide different perspectives on how process mining is used in practice. For instance, [11] assesses the maturity of the field from a practical viewpoint by considering the diffusion of tools and the thoroughness of the application of process mining methodologies over the years. However, it does not cover the specific details of how the questions are answered in the case studies except for an enumeration of 7 process mining techniques used. Other studies focus on a specific field. For instance, [21] discusses healthcare case studies according to 11 main aspects, but the level of abstraction is similar to [11]. A similar analysis applied to the BPIC reports is performed in [16]. The authors focus on the methods, tools, and techniques used in the reports submitted by the participants. None of these papers links the analysis techniques used to the business questions addressed nor discusses the context in which these techniques are applied. The closest work to ours is [14] and [27]. Klinkm¨uller et al. [14] qualitatively analyze BPIC reports to understand how process analysts perform their work. The focus is put on visual representations and their information needs for all types of questions. We complement this research with a narrower but deeper analysis. We focus on identifying all specific low-level operations that are used to answer time performance questions. Because of that, the operations identified in our paper are more fine-grained, which brings a more precise understanding about how questions related to time performance are addressed. Zerbato et al. [27] conduct an empirical study to understand how analysts perform a process mining task. Their study focuses on the initial exploratory phase of process mining where analysts examine and understand an event log. It reveals that the 12 analysts who participated in the research follow different behavior patterns when exploring event logs with Disco, and identifies some typical operations to carry out process mining. We complement this research by focusing on specific business questions and looking beyond the exploratory phase. Specifically, we identify operations related to time performance that have been performed by analysts with different profiles (i.e. various organizations and countries) making use of several process analysis tools (e.g. Disco, ProM and Celonis). Moreover, specific techniques and visualizations have been developed to analyze the time perspective of a business process, its cycle time, and its bottlenecks (e.g. [13,17,20,22,25]). With respect to them, this paper helps to understand if they are used in practice and the context in which they could replace some of the more general techniques used in the BPIC reports. 3 Research Methodology We apply a methodology similar to the one proposed in [19], which follows a mixed-method approach that combines qualitative and quantitative research methods. We first perform a qualitative coding similar to the one proposed in [14]. This coding allows us to quantitatively analyze the operations in the BPIC reports. With this study we want to answer 4 research questions: RQ1: What operations are used to answer the questions on time performance? We aim to identify the analysis operations frequently used in time performance analysis. RQ2: What types of answers to time performance questions can be identified? We aim to discover categories in the answers provided by the authors of the reports depending on the operations used. This can inspire future process analysts. RQ3: How does the context affect the similarity of the answers to time performance questions? The context involves the specific goal pursued in the question, the event log analyzed and the authors’ profile. We aim to find commonalities and differences regarding these aspects to understand how they affect the answers. RQ4: What is the observable impact of the current state of the art on the answers to time performance questions? We aim to understand how the existing tool support and literature help and limit the answering of the questions. The steps of the methodology are described next. Details and materials are available at our repository [5]. 3.1 Step 1: Data Collection As we focus on questions related to time performance in business processes, the first step was to review which BPIC had business questions concerning this perspective. We found a total of 7 questions related to time in 4 editions: 2015 [6], 2017 [7], 2019 [8] and 2020 [9]. In BPIC 2020, we noticed that many operations and data of the first question were reused to answer the second one, so we have considered them as the same question. We classified these questions attending to their goal in: differences, whose goal is to find differences between the throughput Table 1. Questions related to time performance in the BPI challenges since 2011. The column of BPIC (ID) represents the selected questions and the identifiers of the questions used to refer them (e.g. 2015-Q5 whose identifier is C15, represents question 5 of challenge 2015). BPIC (ID) Type Question Answers 2015-Q5 (C15) Differences Where are differences in throughput times between the municipalities and how can these be explained? 9 2017-Q1 (C17) Differences, fragments What are the throughput times per part of the process, in particular the difference between the time spent in the company’s systems waiting for processing by a user and the time spent waiting on input from the applicant as this is currently unclear? 21 2019-Q2 (C19) Fragments What is the throughput of the invoicing process, i.e. the time between goods receipt, invoice receipt and payment (clear invoice)? To answer this, a technique is sought to match these events within a line item, i.e. if there are multiple goods receipt messages and multiple invoices within a line item, how are they related and which belong together? 12 2020-Q1, 2020-Q2 (C20A) Fragments, differences What is the throughput of a travel declaration from submission (or closing) to paying?, Is there are difference in throughput between national and international trips? 20 2020-Q4 (C20B) Fragments What is the throughput in each of the process steps, i.e. the submission, judgement by various responsible roles and payment? 17 2020-Q5 (C20C) Bottlenecks Where are the bottlenecks in the process of a travel declaration? 18 2020-Q6 (C20D) Bottlenecks Where are the bottlenecks in the process of a travel permit (note that there can be multiple requests for payment and declarations per permit)? 13 Total - - 110 of different processes; fragments, whose goal is to calculate the throughput of parts of the process; and bottlenecks, whose goal is to find bottlenecks. A question can be related to more than one goal (cf. Table 1). We considered only the reports that answer the selected questions, specifically, those that have a specific section dedicated to respond to a question. As a result, 62 reports and 110 different answers were included in the analysis: 9 of 9 reports in 2015, 21 of 24 in 2017, 12 of 15 in 2019, and 20 of 37 in 2020. The number of answers to questions in 2020 varies because not every report provided an answer to every question. Additionally, the reports were grouped according to the authors’ profile: students, professionals, and academics. The distribution of reports and answers in these profiles per year can be found in our repository [5]. 3.2 Step 2: Coding We followed an inductive category development based on several coding iterations. The way in which these iterations were performed was inspired by the Grounded Theory methodology [23]. First, we applied open coding to the answers to the questions provided in the BPIC reports. This involved reading the answers and marking them with annotations to derive codes. During this initial phase, we noted that each report answered the time performance questions using their own specific terms, but these terms referred to the same concepts. Thus, we had Table 2. Examples of coding Text in the report Annotation Operation Variant For the former case, filtering was performed by designating A Pending as forbidden and OCancelled as end activity Filter traces depending on the lack of A Pending and O Cancelled as end point Filter traces Filter traces by activities By filtering all cases that did not have a project number in Disco Filtering of traces without project number in disco Filter traces Filter traces by organizational units to unify the vocabulary to compare the annotations of different answers more easily, since our purpose was to discover commonalities among the answers. To do so, we created a key concept code where we related different terms to a unique concept. For instance, in some cases the authors referred to the total execution time of the process as throughput but the implementations provided calculate the time required to complete (a part of) a process (cycle time), so we decided to rename it to cycle time. The name throughput was kept in the cases where the number of activities or process instances per time unit is calculated. Afterwards, we grouped the annotations of the answers by some time performance questions that we sampled to better handle the annotations depending on their similarity. Once the annotations were grouped, we could identify their corresponding operations and detect the same operations from different reports. For instance, in two different reports of the BPIC 2020 we found the two similar annotations shown in Table 2. In both annotations the authors are filtering traces, despite using different criteria. Therefore, we grouped them into an operation called Filter traces. This way, we created an operation code to avoid defining similar operations with different names. This also made us notice that the implementations of some operations had the same purpose but were performed over different variables. We call them variants. For example, the aforementioned annotations are two variants of how to filter traces, since one is filtering by activities and the other by organizational units. Thus, we labelled them as variants of Filter traces as depicted in Table 2. The coding process was iterative and finished when no more operations and variants were obtained from the reports. In total, we found 55 operations with 137 variants. To mitigate the bias of one researcher having to identify the codes, during the whole process two authors annotated and coded a subset of the reports independently and then shared the results. In case of disagreement, the four authors discussed the differences to reach a consensus. Moreover, we categorized the operation codes into 6 types based on their goals as explained in Sect. 4.1. 3.3 Steps 3 and 4: Dataset Creation and Quantitative Analysis Next, we handcrafted a dataset that relates the operations and variants identified to the answers in which they appear. We also included metadata related to the question, year, category, and type. The resulting dataset has 955 actions and 110 answers, where an action is an execution of an operation variant. Finally, we performed a quantitative analysis of the dataset to answer the research questions. Specifically, we analyzed the dataset using frequency distributions and descriptive statistics to answer RQ1. In order to respond to RQ2 we analyzed the answers depending on the number of performed operations and we applied KMeans clustering of the answers in the reports according to the operations used in them. To answer RQ3, we used the Sørensen- Dice coefficient [24] between pairs of answers to find similarities. This index DSC(A, B)=(2|A∩B|)/(|A|+|B|) measures the similarity between two sets Aand B, where 0 indicates two totally different sets and 1 two equal sets. Furthermore, we used 45 of the 72 measures described in [26] to retrieve properties about the event logs that could help us to cluster them to understand how they can affect the answers. We excluded those measures that had problems during their computation (e.g., too long execution times). Finally, to answer RQ4, we checked the existing tool support in process mining tools as well as related literature. More details of this step are provided in Sect. 4. The codes in bold in the first column of Table 1will be used therein for the sake of brevity. 3.4 Threats and Limitations First, as it often happens in qualitative research, there could be personal bias because the annotations and coding rely on a subjective interpretation of the description that appears in the report. We mitigate this threat as discussed above, but a residual risk remains. Second, the conclusions of this study are based on reports that address 8 time performance questions. These questions deal with typical temporal problems, such as bottlenecks or differences in throughput time between processes. Although the sample is representative, it does not cover all possible questions. Third, the reports analyzed could be done by the same organization and hence, be more similar to each other than otherwise. We checked the organizations of the reports and found that there was a predominant organization with 10 reports of 62. However, the similarity using the DSC index between the reports of this organization is smaller than the similarity with the reports belonging to participants from other organizations (0.12 and 0.16, respectively). Thus, we concluded that including them would not bias the results. Finally, the analysis is based only on the reported answers. This has two implications. First, the order in which the operations appear in the report may be different from the order in which the analysts performed them. To partially mitigate this risk, we ignore the operations order for our analysis. Second, not all operations used by the analysts may appear in the report. Some of them might not give relevant results and be omitted in the report, and others might have been removed because of space restrictions. This risk cannot be fully avoided with our study design. However, we can safely assume that the operations that appear in the report are those that the authors found more relevant. Furthermore, as long as one of the operations appears in one answer, it is considered in our analysis. Table 3. Classification of operations sorted in descending order of frequency. In bold are those with a frequency higher than the average (17.05). Operation (absolute frequency - number of variants - number of questions) OPERATIONS TO ANALYZE TIME: Calculate cycle time (152-12 - 7),Find bottlenecks (63 - 5 - 6),Compare cycle time (30 - 1 - 7),Calculate waiting time (27 - 1 - 2),Calculate throughput (18 - 1 - 3), Calculate processing time (10 - 1 - 3), Compare throughput (2 - 1 - 1), Compare waiting time with processing time (2 - 1 - 1), Analyze cycle time depending on the events (1 - 1 - 1), Calculate intervals of time of the traces (1 - 1 - 1) OPERATIONS TO MANIPULATE THE DATA: Filter traces (86 - 7 - 7),Group traces (58 - 12 - 6), Preprocess the traces of the logs (11 - 1 - 5), Filter events (11 - 4 - 5), Group activities (8 - 4 - 3), Filter activities (9 - 4 - 5), Filter sub-processes (6 - 2 - 1), Filter variants depending on frequency (2 - 1 - 1), Preprocess the events of the logs (2 - 1 - 1), Group events by attributes (1 - 1 - 1), Group events by time (1 - 1 - 1), Group organizational units (1 - 1 - 1), Group sub-processes (1 - 1 - 1) OPERATIONS TO CALCULATE STATISTICS Calculate number of elements (76 - 7 - 7),Calculate percentages (55 - 4 - 6),Calculate statistics (36 - 4 - 6),Calculatefrequency(25-7-7), Calculate average of activities per trace (3 - 1-3) OPERATIONS TO REPRESENT THE PROCESS GRAPHICALLY: Represent process map (47 - 2 - 7),Represent bar charts (36 - 6 - 6),Represent histograms (32 - 3 - 7),Represent temporal series (25 - 4 - 5), Represent heat maps of cycle time and an attribute (6 - 1 - 3), Represent linear tendency of cycle time with respect an attribute (5 - 1 - 2), Represent scatter plot of cycle time and an attribute (5 - 1 - 4), Represent circular charts of attributes of the traces (3 - 1 - 1), Represent box plots of cycle time (3 - 1 - 1), Represent density diagram of cycle time (2 - 1 - 2), Represent lineal distribution of an attribute by traces (2 - 1 - 1), Represent correlation graph of variables (1 - 1 - 1) OPERATIONS TO IDENTIFY ELEMENTS IN THE DATA: Identify attributes (34 - 3 - 6), Identify resources (10 - 3 - 3), Identify transitions by cycle time (10 - 1 - 4), Identify organizational units (9 - 3 - 1), Identify activities (8 - 4 - 4), Identify roles (7 - 2 - 3), Identify traces by cycle time (2 - 1 - 2), Identify specific sub-processes (1 - 1 - 1), Identify impact of bottlenecks by organizational unit (1 - 1 - 1) OTHERS: Calculate dates of the development of activities of resources (2 - 1 - 1), Assign resource to each activity (1 - 1 - 1), Apply techniques of machine learning (1 - 1 - 1), Apply decision trees (1 - 1 - 1), Discover happy path of the process (1 - 1 - 1), Discover process maps (1 - 1 - 1) Despite these limitations, we believe that the use of the BPIC reports also brings relevant advantages as discussed earlier. Furthermore, we think that the analysis conducted provides relevant insights that can be used as a starting point to improve our understanding of how questions are answered in practice. 4 Results Next, we describe how we have addressed the research questions defined in Sect. 3 and the results obtained. 4.1 RQ1: Operations Used to Answer Time Performance Questions We identified 55 different operations and 137 variants and classified them in 6 groups according to their purpose (cf. Table 3). The operations that do not fit in any of these groups are classified as others. Table 3also shows the absolute frequency of each operation, the number of variants identified for each operation and the number of questions for which at least one answer uses each operation. Most of the operations (35) have only one variant. The others have between 2 and 7 variants, except Group traces and Calculate cycle time (CT) with 12 variants each. In the following, we outline how operation variants are defined. The operations to analyze time focus on the temporal analysis of the process, such as calculating and comparing cycle time and waiting time, or finding bottlenecks. Two operations have more than one variant. Calculate CT can be implemented for different process elements (e.g. the whole process or pairs of events) and considering either all traces or subsets of them. Find bottlenecks varies depending on where to look for the source of the bottleneck (e.g. activities or process fragments) and the criteria used to consider that a bottleneck is happening (e.g. activities that exceed the average cycle time of all activities). The operations to manipulate data reorganize the traces or the events from the log, including their filtering, grouping, and preprocessing. Concerning their variants, filters and groupings are applied on some process element (e.g., traces or activities) and implemented depending on a condition related to a temporal performance measure, an attribute of the event log, or another process perspective. For instance, traces can be filtered depending on the existence of activities and activities can be grouped according to certain thresholds of cycle time. The operations to calculate statistics give numerical insights applying descriptive statistics, such as counts, proportions and frequencies. Regarding their variants, Calculate number of elements,Calculate percentages and Calculate frequency are implemented depending on the process element to which they apply, like calculating the number, percentage or frequency of each activity, or to calculate the number of values that an attribute takes or the frequency with which one of them occurs. As for Calculate statistics, this operation is applied to cycle time, throughput and activities. The operations to represent the process graphically show visual insights of the process by creating process maps, bar charts, or histograms among others. Their variants are based on what is being represented (e.g. cycle time), or the values of some attribute (e.g. Represent process map with CT). Additionally, Represent temporal series and Represent bar charts vary depending on the process element. Temporal series are also used to represent throughput. The operations to identify elements in the data find a specific aspect of the process and its context, such as process fragments, activities, attributes (e.g. roles and resources) based on some condition. In this case, the variants represent the conditions used to identify the elements (e.g. Identify attributes by CT). 4.2 RQ2: Types of Answers We have analyzed the answers from two different perspectives. First, we have compared the average number of total and distinct operations per questions, which are collected in Table 4. We believe that there can be two factors related to the difference between questions: (i) the question itself, e.g. C19 is broader and hence, requires more operations to give a proper answer; and (ii) the fact 4. Cabanillas, C., Ackermann, L., Sch¨onig, S., Sturm, C., Mendling, J.: The RALph miner for automated discovery and verification of resource-aware process models. Softw. Syst. Model. 19(6), 1415–1441 (2020). https://doi.org/10.1007/s10270-020- 00820-7 5. Capit´an-Agudo, C., Salas-Urbano, M., Cabanillas, C., Resinas, M.: BPI challenge analysis: how are time performance questions answered, March 2022. https:// github.com/isa-group/bpi-challenge-performance-analysis 6. van Dongen, B.: BPI Challenge 2015. 4TU.ResearchData, May 2015. https://doi. org/10.4121/uuid:31a308ef-c844-48da-948c-305d167a0ec1 7. van Dongen, B.: BPI Challenge 2017. 4TU.ResearchData, February 2017. https:// doi.org/10.4121/uuid:5f3067df-f10b-45da-b98b-86ae4c7a310b 8. van Dongen, B.: BPI Challenge 2019. 4TU.ResearchData, January 2019. https:// doi.org/10.4121/uuid:d06aff4b-79f0-45e6-8ec8-e19730c248f1 9. van Dongen, B.: BPI Challenge 2020. 4TU.ResearchData, March 2020. https:// doi.org/10.4121/uuid:52fb97d4-4588-43c9-9d04-3604d4613b51 10. van Eck, M.L., Lu, X., Leemans, S.J.J., van der Aalst, W.M.: PM2:aprocess mining project methodology. In: CAiSE, pp. 297–313 (2015) 11. Emamjome, F., Andrews, R., ter Hofstede, A.H.: A Case Study Lens on Process Mining in Practice. In: OTM Conferences. pp. 127–145 (2019) 12. Graafmans, T., Turetken, O., Poppelaars, H., Fahland, D.: Process mining for six sigma. Bus. Inf. Syst. Eng. 63(3), 277–300 (2021) 13. Hompes, B.F.A., Maaradji, A., Rosa, M.L., Dumas, M., Buijs, J.C.A.M., Aalst, W.M.P.v.d.: Discovering causal factors explaining business process performance variation. In: CAiSE, pp. 177–192 (2017) 14. Klinkm¨uller, C., M¨uller, R., Weber, I.: Mining process mining practices: an exploratory characterization of information needs in process analytics. In: BPM, pp. 322–337 (2019) 15. de Leoni, M., van der Aalst, W.M.P., Dees, M.: A general process mining framework for correlating, predicting and clustering dynamic behavior based on event logs. Inf. Syst. 56, 235–257 (2016) 16. Lopes, I.F., Ferreira, D.R.: A survey of process mining competitions: the BPI challenges 2011–2018. In: BPM Workshops, pp. 263–274 (2019) 17. Low, W.Z., van der Aalst, W.M.P., ter Hofstede, A.H.M., Wynn, M.T., De Weerdt, J.: Change visualisation: analysing the resource and timing differences between two event logs. Inf. Syst. 65(Supplement C), 106–123 (2017) 18. Maggi, F.M.: Discovering metric temporal business constraints from event logs. In: Johansson, B., Andersson, B., Holmberg, N. (eds.) BIR 2014. LNBIP, vol. 194, pp. 261–275. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-11370-8 19 19. Revoredo, K., Djurica, D., Mendling, J.: A study into the practice of reporting software engineering experiments. Emp. Softw. Eng. 26(6), 1–50 (2021). https:// doi.org/10.1007/s10664-021-10007-3 20. Richter, F., Seidl, T.: TESSERACT: time-drifts in event streams using series of evolving rolling averages of completion times. In: BPM, pp. 289–305 (2017) 21. Rojas, E., Munoz-Gama, J., Sep´ulveda, M., Capurro, D.: Process mining in healthcare: A literature review. J. Biomed. Inform. 61, 224–236 (2016) 22. Senderovich, A., et al.: Conformance checking and performance improvement in scheduled processes: a queueing-network perspective. Inf. Syst. 62, 185–206 (2016) 23. Stol, K., Ralph, P., Fitzgerald, B.: Grounded theory in software engineering research: a critical review and guidelines. In: ICSE, pp. 120–131 (2016) 24. Sørensen, T.: A method of establishing groups of equal amplitude in plant sociology based on similarity of species content and its application to analyses of the vegetation on Danish commons. Am. J. Plant Sci. 5, 1–34 (1948) 25. Wynn, M.T., et al.: ProcessProfiler3D: a visualisation framework for log-based process performance comparison. Decis. Support Syst. 100(Supplement C), 93– 108 (2017) 26. Zandkarimi, F., Decker, P., Rehse, J.R.: Fig4PM: a library for calculating event log measures. In: ICPM Doctoral Consortium and Demo Track, pp. 27–28 (2021) 27. Zerbato, F., Soffer, P., Weber, B.: Initial insights into exploratory process mining practices. In: BPM Forum, pp. 145–161 (2021)