Enhancing Personalized Learning Through Process Mining

Since the 1980s, the application use cases and relevance of data mining in business have substantially increased (Chen et al. 2012). We follow the definition of Bissantz and Hagedorn (2009, p. 118), who define data mining as “the extraction of implicitly available, non-trivial, and useful knowledge from large, dynamic databases with relatively complex structures.” By facilitating better monitoring, identification, and analysis of raw business data, data mining aims to automate data preparation and analysis (Bissantz and Hagedorn 2009). Process mining intends to help organizations improve their understanding of processes by means of the identification of variants of a business process, identification of noncompliant behavior, performance of relevant insights, or assurance of the quality of discovered process model (van der Aalst et al. 2012, vom Brocke et al. 2021).

The application of data mining in the educational domain referred to as educational data mining evolved as a result of the availability of educational data collected by TML environments (Lemay et al. 2021). A related concept is learning analytics (see Fig. 1 for an overview and proposed demarcation of related concepts). Learning analytics aims to discover useful insights into a user’s learning behavior from large educational data sets and activities, that is, the discovery of information (i.e., what is the current sentiment of the learner), generation of learning progress (i.e., what is the learner’s relative and absolute progress) or understanding learner behavior (i.e., what are the personality and learning routines like) (Nguyen et al. 2020a). Hence, learning analytics embodies a learner-centered perspective (e.g., employ analytics to inform or empower instructors or learners about learner gaps or preferences). Learning analytics applications focus on evoking data-based interventions for learners to enhance their learning behavior (Chatti et al. 2012). Methods of learning analytics include classifications and predictions to monitor, individualize, and enhance learning behavior, the visualization of learning data and insights, social network analysis to identify relationships between learners, and optimization of individual learning activities (Chatti et al. 2012). Regarding the data sources, learning analytics can process various large data sets that are generated by TML environments and also other sources, such as social networks or physiological data.

Educational process mining can be viewed as a subfield of learning analytics, which uses process mining techniques to add a process-centered perspective on learning behavior (Bogarín et al. 2018; Juhaňák et al. 2019; Cerezo et al. 2020). More specifically, educational process mining aims to analyze log data from TML environments to monitor, discover, and enhance learning processes (Cerezo et al. 2020). Its unique characteristic is that it specifically deals with the analysis of educational data logs and, therefore, aims to better understand learning processes, that is, learning process discovery, social network discovery, process monitoring, process analysis, or the enhancement of processes and flows. Instead of analyzing single data points of the learner’s journey, educational process mining’s process-oriented perspective assists in uncovering the entire end-to-end learning process and the dependencies among the activities in the process, that is, the learner’s success or other measures of success.

In the following, we refer to the three major educational process mining activities of discovering, monitoring, and enhancing learning processes and activities. Consequently, educational process mining uses analytical methods for educational data. The academic discourse around educational process mining is increasing because this subfield of learning analytics offers a new, holistic, process-centered perspective on learning. Such a dynamic perspective is beneficial for individuals, educators, and organizations alike. The focus on learner processes during their educational development is a pedagogically beneficial perspective and has been well described in literature (Roth 1970; Söllner et al. 2018). Thus, educational process mining is a technology that could bring back the learning-process-centered perspective developed by Roth (1970) and thus offers a process-centered perspective in the analysis of learning paths. Figure 2 illustrates an architecture of relevant scenarios in practice.

User data is generated automatically as a learner interacts with the learning environment (such as the system in Fig. 2) through clicks, chat protocols, review comments, or specific learner content, such as written student texts. These event logs are then used to fuel the learning process model, which is used to discover, to conform, or enhance the system. With the continued expansion of TML, an enormous amount of student-centered data can enrich the pedagogical embedding itself (e.g., Wambsganss et al. 2021) through personalization, recommendations (Nguyen et al. 2020b), and formative feedback (Hattie and Timperley 2007). As well it can also support the educators and institutions who monitor and evaluate learning processes by extracting information from the models and acting on the findings.

Three basic types of process mining can be distinguished (van der Aalst 2012; Bogarín et al. 2018):

Data mining and process mining, and more specifically, their subfields of learning analytics and educational process mining methods, can be used in the educational domain to enrich learning scenarios with data-driven value. However, this requires an understanding of the anatomy of TML (a general overview can be found in Gupta and Bostrom 2009). With their call for TML research 20 years ago, Alavi and Leidner (2001) defined TML as “an environment in which the learner’s interactions with learning materials, peers, and/or instructors are mediated through advanced information technology” (Alavi and Leidner 2001, p.2). As Gupta and Bostrom (2009) noted, TML includes “all the elements of a social-technical system: technology and learning techniques, process, actors, actions, and outcomes” (Gupta and Bostrom 2009, p.3). Scholars in different disciplines have used the terms e-learning, online learning, distance learning, technology-enhanced learning, or IT-supported learning synonymously (Gupta and Bostrom 2009). TML is based on the underlying assumptions that a particular learning context influences the embedded learning structures (DeSanctis and Poole 1994).

In this paper, we follow a constructivist understanding of learning, which describes how knowledge is always constructed by the learner and can be influenced by experiences (Vygotsky 1980). In this vein, the learning experience is influenced by a learning method. The framework of TML focuses on the learning process of a learning activity. By definition, a learning process describes the sum of the individual activities a student performs. According to Gupta and Bostrom (2009), it describes how a student interacts with and adapts the learning method structures. A learner’s process is a complex phenomenon and includes variables such as cognitive processes and interactions related to the learning methods. Gupta and Bostrom (2009) define learning outcomes as the pivotal measurement of the success of a learning process in TML. Accordingly, learning outcomes are “the target assessment or measures used to determine the achievement of learning goals” (Gupta and Bostrom 2009, p. 691), e.g., in a certain individual learning scenario. With “individual learning scenario”, we mean a certain pedagogical set-up where a learner receives adaptive feedback on her own learning progress, for example, from a human educator (or artificial tutoring system) that recognizes learner gaps, anticipates needs, suggests suitable learning materials, and provides individualized feedback.

Educational process mining can contribute to different elements of the TML framework. It can have a structural impact (e.g., an impact on learning method structures as a technology that directly influences the learning process of students) or a process impact (e.g., by using it to better measure appropriation or learning outcomes).

Literature on TML has mostly focused on the implementation and analysis of personal and adaptive learning environments, e.g., based on adaptive formative feedback (Hattie and Timperley 2007; Ifenthaler and Gibson 2019). Here, perspectives in learning analytics, such as educational process mining, can enhance the adaptivity or the possibility of individual feedback and recommendations in TML environments (Bogarín et al. 2018). Nevertheless, literature has in the past mostly approached the use of process mining for educational purposes from a technical perspective. For example, Mouchel et al. (2023) used process mining to better understand the learning and revision behavior of students who received intelligent writing support (Mouchel et al. 2023). Ludwig et al. (2024) applied a sequence mining approach to behavioral process data to predict problem-solving success of students, which ultimately allows instructors to better intervene with personalized scaffolding. Another example is He et al. (2021), who applied different process mining techniques, such as Fuzzy miner or pMiner, to 122,167 event logs from 527 undergraduate students extracted from the learning management system Moodle in order to track students’ self-regulated learning patterns in response to a formative assessment (He et al. 2021). Also, research in information systems and educational technology has motivated research and studied the effect of various learning scenarios and configurations based on educational process mining (e.g., Johnson et al. 2019; Cerezo et al. 2020). This includes, for example, the investigation of the impact of metacognitive prompts on self-regulated learning (Engelmann and Bannert 2019), the investigation of student adherence to a recommended course path (Cameranesi et al. 2017), and the investigation of process-based feedback during medical training (Lira et al. 2019). The expanding number of interdisciplinary studies on educational process mining emphasizes the importance of a holistic understanding of its design, capabilities, and potential repercussions (Bogarín et al. 2018; Costa et al. 2020).

Still, research is distributed over a wide range of socio-technical and technological viewpoints, resulting in a critical lack of an integrated perspective. For instance, in their systematic literature review, Ghazal et al. (2018) did not derive specific educational process mining traits and dimensions and instead concentrated on technical factors. Bogarín et al. (2018) have presented a systematic summary; however, they did not use a clear methodological approach, such as a systematic literature review (Webster and Watson 2002; vom Brocke et al. 2015). Costa et al. (2020) only considered process mining literature on Moodle, making the review inapplicable to other learning scenarios. Assessing the potential of educational process mining for individualized learning scenarios and deriving suitable design characteristics requires an integrative perspective based on TML and the potentials described before. Here, classifications of educational process mining are beneficial for developing tailor-made use case applications for different stakeholders (e.g., educators, institutions, or learners). Several scholars have highlighted the lack of shared information about the embedding of process mining in various TML scenarios, as well as the lack of insights into the involved users (e.g., educator, institution, or learner), the technological context (e.g., the discovery algorithm used), or the pedagogical structure (e.g., in which educational domain or for which pedagogical task process mining is analyzed (Ghazal et al. 2018; Rogiers et al. 2020). In fact, present reviews fall far short of providing a complete and solid structure for educational process mining applications. A systematic classification of educational process mining scenarios from a TML would allow researchers to more effectively design, evaluate, compare, and theorize how different technological embeddings of the young field of process mining impact student learning outcomes in a specific learning scenario and task. A TML perspective would allow the classification of a certain information system into key aspects (user, task, structure, and technological standpoint), which results in varied configurations and outputs (Bostrom and Heinen 1977; Gupta and Bostrom 2009). We want to fill this knowledge gap by deriving a unique taxonomy that aids decision-making when developing, designing, and evaluating relevant scenarios and applications, as well as by specifying the links of educational process mining characteristics to the outcome of a learning scenario. Educational process mining can serve as a key for creating and establishing individualized learning scenarios and help information systems researchers and educators design and analyze digital learning processes more effectively. Here, the design and analysis of learning processes refers to the systematic creation and evaluation of educational strategies and methods to enhance student learning experiences and outcomes. This involves developing educational interventions (design) and assessing their effectiveness and impact on learning (analysis).

We conducted a literature review to identify relevant material as the foundation for the systematic construction of a taxonomy according to Webster and Watson (2002), vom Brocke et al. (2015), and PRISMA from Mother et al. (2009). Based on recent literature reviews on educational process mining (e.g., Bogarín et al. 2018; Costa et al. 2020), we identified different keywords that researchers used to describe process mining in the educational domain.Based on the keywords, we created the following search string: [“Process Mining” OR “Workflow Mining” OR “Task Mining”] AND [“Education” OR “Learning Analytics” OR “Training” OR “Skill Development” OR “Student” OR “Teaching” OR “Learner” OR “Pedagogic” OR “University”]. We selected two major areas for educational process mining research: information systems and educational technology. They cover a sizable portion of the literature on our topic of interest. To find relevant studies that applied process mining in educational scenarios, we applied the search strings to the following six databases: AISeL, EBSCO, ScienceDirect, ProQuest ABI Inform, IEEE Xplore, and ACM Digital Library (see Fig. 4).

The process of applying the four steps of PRISMA (Mother et al. 2009), namely Identification, Screening, Eligibility, and Inclusion, is presented in Fig. 4, along with justification for the inclusion and exclusion of studies in each step.

We identified 3028 studies from six databases in January 2024. After the initial screening of titles, abstracts, and keywords and removing of duplicates, we counted N = 2944 studies. Next, we reviewed the studies to ensure that they met the selection criteria and applied to the subject of our study. We eliminated studies that did not mention process mining or that used process mining in an area other than education. Many manuscripts were rejected because they represented a distinct study scope. Several studies, for example, addressed mineral mining or machine learning and process mining outside of an educational environment and were thus excluded from our sample. This filtering yielded 77 empirical studies for review that referenced using process mining in education throughout their research. Following that, a forward and backward search was performed following Webster and Watson (2002). Two publications were added to the list after checking the references, resulting in 79 relevant papers.

Our goal was to offer a holistic yet nuanced framework on educational process mining. We used the taxonomy development method of Nickerson et al. (2013), which has been used in several previous information systems studies (e.g., Singh and Varshney 2020). Their approach to constructing taxonomies is iterative and systematic and is based on theoretical underpinnings (deduction) and empirical data (induction). We constructed several dimensions and characteristics based on the published research regarding process mining in education and the empirical evidence of certain meta-attributes. The purpose of a meta-characteristic is to systematically identify design and analysis elements for a digital learning process based on educational process mining applications to better identify, design, and compare process mining applications in educational contexts. To do so, we looked at educational process mining design characteristics from a TML perspective (Gupta and Bostrom 2009) to form a holistic contribution to the current knowledge of process mining in the specific domain of education.

To assure the quality of our taxonomy, we evaluated it using the five listed criteria: comprehensibility, conciseness, robustness, extendibility, and explanatory power (Nickerson et al. 2013). We conducted semi-structured interviews with three experts in the analysis of digital learning processes with educational process mining and three experts in the design of TML environments following the taxonomy evaluation guidelines of Szopinski et al. (2019) (see Table 2). Both groups represent experts in the domain and potential end users of our taxonomy. The shortest interview was 30 min, while the longest lasted 85 min. The interview guideline included 18 open-ended questions based on the five listed assessment criteria. Before the interviews, we emailed the final copies of our taxonomy, the meta-characteristic, and exemplary process mining scenarios to the interviewees. We asked interviewees to remark on and provide ideas for modifying and refining the taxonomy during the interviews. We redesigned the taxonomy after obtaining some ideas for dimension categorization and other minor improvements from the experts, which resulted in the final taxonomy. Following the expert interviews, several significant changes were made to the taxonomy. Originally, the taxonomy dimensions were categorized as T4 (see Table 1): Task/Structure, User, and Technology. A majority of interviewees gave us detailed feedback on the explanatory power of our dimensions and characteristics. For example, interviewees 1–5 suggested using more self-explaining terms for educators and researchers to understand the overall structure of the taxonomy given our meta-characteristics. Hence, based on interviewees 3 and 4 suggestions we renamed and reorganized the dimension of Task/ Structure to Purpose to encapsulate a wider scope, including the application focus and learning mode. This change reflects the feedback that the taxonomy should capture the broader objectives and instructional methods involved in educational process mining. Or, based on the suggestion of interviewees 1 and 2, we split the original Technology dimension into Data and Analysis. This change was driven by the need to highlight the importance of data quality and sources more prominently and to better reflect the variety of analytical approaches used in process mining.

These changes were made to ensure the taxonomy is both comprehensive and practically applicable, addressing the identified gap between theory and practice. This redesigned taxonomy, informed by expert insights, better captures the complexities and varied aspects of educational process mining.

We conducted a cluster analysis on the 79 systematically identified studies on learning process design and analysis (Kaufman and Rousseeuw 2005). Our goal was to find clusters to better comprehend the use of certain educational process mining application groups. We constructed a binary data matrix indicating the presence (1) or absence (0) of specific educational process mining characteristics identified from literature across different studies. These characteristics, aligned with our taxonomy’s dimensions, and the studies, ordered alphabetically by first author, formed the rows and columns of the matrix, respectively. The database consisted of the final eleven dimensions and 45 characteristics from our final taxonomy (see Sect. 4 and Fig. 5). Using Ward’s algorithm (Ward 1963), known for its efficacy with smaller datasets, we conducted an agglomerative clustering with a Euclidean metric to find natural groupings and patterns in the data. We discovered three natural clusters of learning scenarios based on educational process mining. These clusters lay the empirical foundation for formulating the patterns for designing and analyzing the digital learning process with process mining.

In the fifth step, we conducted semi-structured expert interviews to dig deeper into the solution patterns for recurring problems in designing and analyzing digital learning processes based on educational process mining. Equipped with the resulting dendrogram, the clustered data, and the taxonomy of learning process design and analysis, we interviewed eleven experts in digital learning design analysis. We applied the guidelines of Myers and Newman (2007) for qualitative interviews for rigorous execution and analysis. We conducted expert sampling following Bhattacherjee (2012) to select suitable interview partners. We chose experts who are experienced educators or hold a role in an educational institution (e.g., program coordinator) and conduct research on the design and analysis of TML learning scenarios. Table 3 provides an overview of our interviewees (E1–E11). We conducted 11 interviews from September until October 2022. The interviews lasted between 31 and 57 min.

Our interview guideline consisted of 28 questions, asking the experts about their general experience in TML, including how they plan, design, conduct, monitor, and improve digital learning scenarios. Moreover, we specifically asked questions according to each dimension and characteristic of our taxonomy (without showing our taxonomy) and about the requirements of solutions to recurring problems in the design and analysis of the digital learning process based on our cluster analysis. In this vein, we dug deeper into the nuances of each cluster and derived the design patterns presented in Sect. 5 (see Fig. 6, 7, and 8). The transcriptions of the interviews were evaluated using qualitative content analysis. Hence, the data were coded, and abstract categories were formed. The coding was performed using open coding to form a uniform coding system during the valuation. We derived several topics in each interview and aggregated the most common ones, resulting in five common topics for the design and analysis of the digital learning process. The codes and topics were “recurring challenges” of educational processes mining (e.g., discovering course-taking behavior or predicting outcomes to provide students with ongoing feedback), “influential factors and requirements” to solve the challenges (e.g., data quality or legal requirements), “exemplary success stories” (e.g., tracing learning behavior to prevent student dropouts or modeling student outcomes to provide feedback) and “design and evaluation steps” to implement the solutions (e.g., concrete steps to discover, monitor, and enhance learning scenarios with educational process mining). Based on these codes, we derived three patterns for designing and analyzing the digital learning process that build on our taxonomy and informed based on a field evaluation with eleven experts in the domain of TML. To structure the design patterns, we follow a similar table format of Dickhaut et al. (2022) and Weinert et al. (2021), since this provides a validated structured, and effective way of presenting the design patterns to different use groups. The categories are based on our codes and include “name”, “goal”, “pattern based on taxonomy”, “challenges”, “influential factors & requirements”, “exemplary success stories”, and “design and evaluation steps”.

Dimensions referring to the design options of digital learning processes are concerned with the purpose that a digital learning process should achieve. In this light, we distinguish between application focus, learning mode, learning outcome, and learning task.

The application focus dimension outlines the learning behaviors and preferences that an instructor hopes to reveal and analyze using an educational process mining application. Learner monitoring refers to situations in which an instructor examines a learning process to acquire insights. An example is Uzir et al.’s (2020) use of educational process mining to monitor whether learners comprehend the offered learning techniques. In that sense, educational process mining can offer valuable feedback on learner development, oftentimes on an aggregated group level. Learner evaluation refers to situations in which educational process mining provides information on and assesses individual learning processes (e.g., bad task performance, clarification issues). In that sense, the evaluation looks back on an individual’s past performance or identified issues in the past process to create options for process-based feedback. Lira et al. (2019) provide an example by investigating process-based feedback during medical training. An application of process mining in the domain of learner recommendations is presently not addressed by the educational literature but would define a system that provides learners with actionable suggestions on how to proceed with their learning process (e.g., based on scaffolding, see Winkler et al. 2021). Hence, different from a learner evaluation purpose, a learner recommendation looks forward into the digital learning process before it is completed. The purpose is to provide learning-process-integrated recommendations.

The learning mode dimension outlines the (scope of the) unit of analysis of digital learning processes in more detail. We discovered that application scenarios of digital learning processes that are suitable for educational process mining focus on individual learning, such as when learners study on their own and simply engage with learning resources (Saint et al. 2020) or collaborative learning, which occurs when learners interact, collaborate, and coordinate learning activities with other learners (Schoor and Bannert 2012). This distinction is seen as a component of the learning mode dimension. The learning outcome dimension refers to the knowledge type that should be trained or created while completing various activities of a digital learning process. Therefore, this dimension recognizes that educational process mining is usable for digital learning processes that focus on factual, conceptual, procedural, and metacognitive knowledge. The characteristics are based on Krathwohl (2002), who defined factual (knowledge) as knowledge of terminology and precise information (e.g., Cerezo et al. 2020), conceptual (knowledge) as an understanding of theories and models, and procedural (knowledge) as subject-specific skills and techniques (e.g., Lira et al. 2019). Metacognitive (knowledge) refers to both strategic and cognitive knowledge, such as debating abilities (Wambsganss et al. 2021). The learning task dimension refers to the learning objectives that learners should achieve when following a digital learning process. We employed characteristics influenced by Krathwohl’s (2002) classifications to differentiate among different difficulty levels of a learning assignment in the form of learning objectives. We used Krathwohl’s concept and differentiated between remembering, understanding, applying, analyzing, evaluating, and creating as learning objectives that a learner can attain during a learning process (Krathwohl 2002). We assume that it is easier to capture data for simpler and objectifiable objectives, such as remembering and understanding.

The second set of dimensions refers to design options of digital learning processes and is concerned with the user who will benefit from analyzing a digital learning process. Despite the field’s infancy and the difficulty in precisely distinguishing and identifying the end user for each case at hand, three primary end users may be established. In this light, we consider the intended main end user and the learning context in which the learning process is embedded.

We opted to separate scenarios by the intended main end user, who will have the most insight into the information that educational process mining will provide. A learner would, in most circumstances, be the one to (1) generate the data of the learning process to be analyzed, (2) receive individualized reports and insights, and (3) use the data analysis output for adjusting their learning strategy. Having a student as the intended primary end user implies that the learner may utilize the insights to enhance their learning or course-taking process by receiving insights on their study input quality, learning performance, or study patterns over time (e.g., Cameranesi et al. 2017). The instructional designer (educator) characteristic refers to an end user who primarily benefits from the insights of analyzed learner data. Instructional designers may use such reports to enhance the instructional design or to intervene in real time to provide students with guided feedback and support them in adapting their learning strategy (e.g., Lira et al. 2019). The primary beneficiary of the organization (learning institution) characteristic is the provider of the educational environment. Learning institutions include universities, MOOC providers, vocational training schools, and continuous training platforms, among others. On the university level, this category of main end users refers to course coordinators, considering learning behavior across courses, years, and individual learners. On an open educational level, it refers to platform providers of MOOC courses. Process mining can be used to analyze MOOCs to reduce dropout rates, for example (e.g., Rizvi et al. 2018). Analyzing learner data in terms of learner course completion, fulfillment, and dropout behavior has the potential to support such users with valuable insights for their course-taking behavior. For this type of main user, analyzing data on an aggregated level (as opposed to the individual learner or course level) is most suitable and insightful.

Finally, the learning context explains the learning environment in which digital learning processes will be deployed. We distinguished between kindergarten–high school (e.g., Gomez et al. 2021), higher education (Engelmann and Bannert 2019), and continuous education (Ariouat et al. 2016), which includes workforce training and programs for personal improvement and vocational training. The learning context can have important implications regarding the type of data available, as well as the legal and ethical treatment of data (i.e., using minors’ data).

The third set of dimensions refers to analysis options of digital learning processes and is concerned with the data dimension (i.e., data input and data collection interface). This input dimension refers to the origin and characteristics of the to-be-analyzed data.

The data input characteristics refer to data prerequisites that need to be created as the basis for further educational process mining-based analysis. Although most applications that analyze digital learning processes use pre-existing event data, such as automatically collecting data through a system, we discovered other data, such as audio (Nguyen et al. 2021), video (Lira et al. 2019) or text (Mittal and Sureka 2014), from which event data can be manually or automatically derived or supplemented. Image data might potentially be used, although no studies presently use it.

The data-collecting interface specifies the tools used to collect the event data. Most applications that aim to analyze digital learning processes collect event data by using an internal system (Juhaňák et al. 2019) or a MOOC platform (Rizvi et al. 2018), other web-enabled learning tools (i.e., accessible and distributed through a web browser from any kind of device), non-web-enabled tools (software accessed as an application on specific devices, e.g., Doleck et al. 2016) and automatic and manual data coding, for example, through the coding of video data, describing the remaining cases.

The fourth set of dimensions refers to analysis options of digital learning processes and is concerned with analysis techniques and the analysis output format. This dimension differentiates between process mining type, analysis beyond process mining, and output presentation. It covers the methods used to examine digital learning processes.

Although various techniques are possible in this area, we opted to focus on distinguishing between the fundamental process mining types employed and the type of analysis that goes beyond the standard process mining functions. Only applications that expressly indicated the usage of discovery and conformance were found in terms of process mining categories. Discovery employs process mining to build a process model by analyzing event log data. An example of this would be creating a process model of a learning process using an event log from a system (e.g., Rogiers et al. 2020). We discovered conformance in scenarios in which process mining is used to compare the mined model based on the event log to an existing process model. One example is checking adherence to course order suggestions (e.g., Cameranesi et al. 2017). Process enhancement was not expressly employed in any of the evaluated articles, yet its applications are conceivable, such as extending a learner’s learning process model using information extracted from specific event logs.

Analysis beyond process mining refers to the many sorts of analyses that are done to either precede or extend typical process mining analysis. This excludes supplemental analyses, which are, in essence, independent of the process mining application. Clustering may be used with process mining to separate learners based on characteristics such as grades. Approaches based on rules can be utilized similarly. Furthermore, the classification may be used to forecast learner success based on previously mined processes. Other approaches resemble the combination of unsupervised and supervised learning or time-series analysis. None entails applications that did not employ analysis beyond process mining techniques.

Finally, the output presentation dimension specifies how the process mining findings are displayed to the intended main end user. The fundamental issue here is that nonexpert users require a greater level of abstraction of information to derive useful insights from the data. Beyond the identified model using process mining, the raw model implies no other form of presentation. A graphical presentation defines how findings are shown, for example, by graphs. In numerical presentation, results are displayed in the form of numbers or tables, such as the fitness scores of identified models. Finally, textual presentation outlines instances in which the system converts the knowledge obtained from process mining into readable information, such as suggestions or automatically created reports.

Discovering student course-taking behaviors in higher education or at MOOC platforms is a common challenge for educational institutions (e.g., Nguyen et al. 2021, 2020a; Cameranesi et al. 2017). A rich analysis of existing course-taking data offers opportunities to identify unknown course combination patterns and provide “next course choice recommendations” to students (Cameranesi et al. 2017). However, due to the growing complexity of multiple available courses, large numbers of students, and countless course combinations, traditional data analysis is not straightforward (Rizvi et al. 2018; Cameranesi et al. 2017). The use of educational process mining to discover and visualize course combinations based on event log data from taken, finished, or dropout courses provides a solution suggestion for this common problem. In our interviewee subset, two interviewees were experts who dealt with course-taking behavior at an institutional level (E10, E11 see Table 3). Both interviewees mentioned that use cases of analyzing course-taking behavior with educational process mining provide multiple opportunities, such as revealing blueprints of commonly taken course combinations to design new profile areas or study programs for students, uncovering course-taking paths that often lead to successful graduation, identifying course combinations that lead to dropouts, or using the discovered information to provide students with “next course recommendations.” For example, E10 mentioned, “we try to collect as much data on the [course] bidding and registration process of students to collect, visualize, and understand the different course-taking behavior of our students.”

Also, in our sample of studies, one natural grouping of studies focused on analyzing course-taking behavior (15 studies, 22% in our cluster analysis). Rizvi et al. (2018), for example, utilized process mining to reduce dropout rates, and Cameranesi et al. (2017) analyzed course-taking behavior to find bottlenecks in a degree program. The pattern to solve this recurring problem for educators and educational designers is to extract event logs about students’ planned, taken, and finished courses from learning management systems. Past researchers have used available process mining tools (majority of studies in our sample use either Disco, ProM, or pMinerR) to apply common discovery algorithms such as alpha miner or heuristic miner to unveil the course-taking behavior of students (Rizvi et al. 2018; Cameranesi et al. 2017). Afterward, the visualization of the individual course-taking paths of students as a graph can be used to monitor the actual behavior. As revealed in our expert interviews (e.g., E5, E6, and E7) and the empirical literature (e.g., Uzir et al. 2020), pattern one involves the application of clustering methods (e.g., hierarchical clustering) to identify novel combinations or blueprints of student course behaviors. Moreover, time series analyses are commonly applied to provide “next course recommendations” to students. For example, E5 mentioned: “We mostly use clustering or time series analysis to track student behavior.” Or E6 mentioned, “we use t-tests between groups and potentially also machine learning algorithms to provide students recommendations after.” Also in the literature, the study of Cameranesi et al. (2017) provides a successful use case application of our pattern. They monitored student course-taking behaviors with process mining discovery and conformance checking based on event logs to investigate if and how students take courses and what might lead to best practices (successful career) or worst-case practices (dropout). Figure 6 summarizes this first design pattern on discovering course-taking behavior at the institutional level.

One of the larger challenges in TML lies in the missing insights on student learning activities due to large-scale or distance learning scenarios and the limited resources of educators to analyze individual learning processes at scale and provide students with needed formative feedback (e.g., Eom et al. 2006; Brinton et al. 2015; Hone and El Said 2016; Huang et al. 2021). Also, almost all of our expert interviewees mentioned that monitoring individual student learning activities to provide formative feedback in the learning processes is still a challenge in most digital learning scenarios (Hattie and Timperley 2007).

Educational process mining can solve this challenge by monitoring and analyzing learning processes. In our cluster analysis, a natural cluster of 20 studies (31%) used process mining to monitor student learning outcomes. Hence, the second design pattern treats the leveraging of interaction data to monitor and enhance the learning activities of students (for an overview, see Fig. 7). This was also a recurrent challenge and application in our interviews with experts as this exemplary quote of E5 depicts: “When we monitor students’ performance and behavior our goal is usually to support them with feedback, e.g., to help students meet deadlines or when students turn in assignments which do not make sense.” Nevertheless, E5 also mentions the limitation of feedback in the same vein: “Psychological problems of students are sometimes hard to support. There is not much you can do. I wish I could help students who struggle emotionally [with process mining and automated feedback].”

A learning process designed with formative feedback based on analysis can enable educators to provide scaffolds and hints to help students in activities they might struggle with, provide personalized feedback, or embed social comparison nudges to trigger students to finish a certain exercise (Hattie and Timperley 2007; Winkler et al. 2021; Wambsganss et al. 2021). Our pattern suggests embedding a learning process in a digital learning tool (e.g., in a quiz, a conversational agent, or a game). It is important that the learning tool leads the student through a sequence of concrete learning activities toward a measurable learning goal. During the learning interaction, the educator can extract event logs of every user interaction (e.g., buttons clicked, text entered, starts and ends of breaks). The goal is to embed the analysis of student learning behavior with discovery and conformance analysis. For example, the Python library PM4PY can be used in the back end of the web tool to provide instant feedback in the form of hints and recommendations if students struggle to find a solution. Wambsganss et al. (2021) provide a successful use case application for pattern 2. They built a conversational agent as a web tool that tutored students through a learning process consisting of a persuasive writing exercise. The authors tracked the students’ interaction data, including the written texts, and provided adaptive formative feedback to argumentative errors. The feedback helped students enhance their argumentative writing skills compared to a control group without adaptive formative feedback.

Formulating learning goals, designing learning activities that address the goals, and organizing both in a curriculum for a particular learning scenario are the main tasks of educators, according to our interviews. One method of monitoring and improving a learning scenario is to monitor individual learning activities to understand and redesign learning processes. A solution to this recurring and common problem could be analyzing student activity event log data to monitor learning activities, evaluate learning goals, and possibly redesign the learning process (Uzir et al. 2020). Almost half of the papers in our sample formed a natural cluster of studies for educational designers in higher education to monitor student learning processes for factual knowledge to understand the learning tasks. Hence, our third identified solution pattern suggests that educators plan a learning process according to learning objectives, user, and purpose (see Fig. 8). Kratzwohl et al. (2002) provide a taxonomy for efficiently setting the right learning goal and learning task. Most of our expert interviewees mentioned that this is their usual practice when starting to plan a learning scenario. For example, E7 mentioned: “We collect everything. The more the better. We collect the click stream with time stamps (log data and trace data). When they watch a video when they stop, etc..) it’s better to have more and filter it. I like doing time series, and pattern analysis, how they perform over time. You can compare them with learning outcomes, motivation, engagement, etc.”

Next, educators need to extract event logs from learning management systems such as Canvas and Moodle and apply common algorithms such as alpha, heuristic, or inductive miner to discover the student learning processes. Through visualizing the learning activities of students as a graph, the educators can then (live) monitor the actual learning processes and compare them with the planned learning process to redesign their curriculum. Uzir et al. (2020) provide a successful use case for applying the third pattern. They used process mining techniques to monitor and cluster log data of 482 undergraduate students over a 13-week course to uncover learning strategies and improve their curriculum. The effort revealed four different learning strategies among the students that were correlated with academic performance. Based on the data analysis, educators established new learning activities for self-regulated learning. However, to be able to utilize the data for learning progress monitoring several interviewees mentioned critical issues about data privacy, ethics, and data quality (E.g., E5, E6, E7). For example, E7 mentioned with regards to learning progress monitoring: “The problem is that the data is not always clean. E.g., the data was not correctly logged. When the data is not clean or actions are missing. E.g. if there is no log-out event, number two is little data (e.g., less than 100 students) this brings additional challenges for us.”

In this study, we conceptualized educational process mining as a new perspective and set of techniques to leverage the potential of learner data in order to improve the individualization of education at scale. This can be done by creating synergies between learning analytics and process mining.

First, our study offers a nuanced understanding of what must be considered when designing and analyzing digital learning processes based on student data in order to discover, monitor, and enhance individual learning based on process mining. We synthesized existing research, including TML, individualization of digital learning, and literature reviews by creating a taxonomy that structured and grouped design characteristics of educational process mining applications (Gupta and Bostrom 2009; Kundisch et al. 2021). Past literature has mostly approached the use of process mining for educational purposes from a technical perspective (e.g., Mouchel et al. 2023, Ludwig et al. 2024, He et al. 2021). Thus, our study contributes to and extends the theoretical foundation of educational process mining by establishing a taxonomy that captures the design characteristics of process mining applications for the specific context of educational scenarios from both a technological and socio-technological viewpoint. Informed by theoretical frameworks in the information systems field (Bostrom and Heinen 1977; Gupta and Bostrom 2009), the taxonomy offers a comprehensive framework to evaluate, design, and compare educational process mining applications in various educational scenarios that become distinguishable through the learning mode or the intended main user, for instance. We uncovered and categorized new dimensions and features that are part of a TML scenario and play a critical role in student learning success beyond the technical perspective of process mining. Specifically, this includes embedding and evaluating algorithmic approaches, as well as incorporating the learner and his or her activities from a pedagogical perspective. These elements are crucial within a pedagogical learning scenario and are instrumental in fostering student learning success. Integrating the TML perspective (Alavi and Leidner 2001), our work offers a holistic view of the role of process mining in education, thereby bridging the gap between technical potential and pedagogical application (Bostrom and Heinen 1977; Gupta and Bostrom 2009).

Second, our study aims to contribute to the understanding of learning analytics and TML in information systems research that could incorporate process mining and improve individual digital learning. Existing literature in the field of TML has largely focused on using digital learning processes for process discovery and conformity approaches (e.g., Rogiers et al. 2020; Cameranesi et al. 2017). Our interview results suggest that current learning scenarios may explain this pattern. Gaps in data traces, for instance, can complicate applying a conformance analysis. Such data traces also provide promising potential for tailor-made individualized course recommendations, well-balanced course allocation for high-quality teaching, or the extraction of new students (Cerezo et al. 2020; Han et al. 2021; Nguyen et al. 2020b).

Third, we used the concept of design patterns to illustrate three use cases of educational process mining. We view patterns as both guidance for data usage and analysis and an illustration of relevant contextual factors to consider (Dickhaut et al. 2023). The three design patterns in this paper suggest data requirements that can enable the realization of respective educational process mining solutions. At the same time, the design patterns also illustrate intended goals and potential challenges to consider in relation to the respective application. We thereby emphasize and challenge the underlying assumptions of effective educational process mining. Understanding and expressing the necessary requirements and potential challenges as well as offering space for solutions is important to build a cumulative body of design knowledge, to show the limits of generalizability, and to prepare the ground for realizing educational process mining applications.

From a practical perspective, our findings on educational process mining enable a more targeted and effective implementation, as well as the analysis and effective use of technology in education. Researchers and practitioners can more effectively design, evaluate, compare, and theorize how different technological embeddings of the young field of process mining impact student learning outcomes in a specific learning scenario and task, thanks to this systematic classification of learning scenarios.

Our taxonomy not only categorizes educational process mining applications but also serves as a practical tool for enhancing the design, delivery, and evaluation of education. It enables educators, administrators, and designers to make informed decisions that directly improve student learning experiences and outcomes. More specifically, the taxonomy aids the identification of patterns and anomalies in student learning behaviors. For instance, by analyzing the process data categorized under different educational scenarios, practitioners can pinpoint areas where students repeatedly encounter difficulties. This identification process is crucial for adapting instructional materials and interventions that directly address these learning gaps. In a similar vein, the taxonomy facilitates the ongoing monitoring of student engagement and progression through their educational activities. By providing a framework to compare students' actual learning paths against optimal process models, educators can intervene in real-time to offer support or adjustments to the course trajectory. This real-time monitoring is particularly beneficial in large-scale learning environments like MOOCs, where individual attention is challenging yet critical for student retention and success.

By enabling the detailed analysis of how different educational processes affect learning outcomes, our taxonomy also guides the design of more effective educational interventions. For example, the taxonomy can help institutions experiment with and refine various teaching methodologies, such as flipped classrooms or blended learning, by providing a structured way to assess their impact on student performance and engagement. The taxonomy thereby supports educational administrators and policy-makers make data-driven decisions about curriculum development, resource allocation, and student support services. By understanding the types of process mining applications that are most effective in various educational contexts, decision-makers can allocate resources more efficiently and develop policies that promote optimal learning environments.

The design patterns of this study are relevant to the educational technology area and associated applications from a practical perspective. The patterns, in combination with the taxonomy for the design and analysis of digital learning processes, serve as a personal guide to studying, designing, and evaluating the individualization of digital learning at scale. We argue that design patterns can provide an actionable space for practitioners to imagine potential use cases of educational process mining for their own scenarios. Particular emphasis is hereby placed on making explicit important underlying assumptions, such as associated data challenges that are fundamental to the effectiveness of educational process mining.

In particular, our three design patterns demonstrate the variety of potential use cases educational process mining offers to TML by improving the individualization of digital learning. Different stakeholders, including instructional designers (design pattern 3), educational organizations (design pattern 1), and individual learners (design pattern 2), can benefit from different interventions. Depending on the user, different types of analyses and output presentations must be considered when deploying process mining for learning analytics. Across the three design patterns, educational process mining offers great opportunities to the higher education domain and individual learning mode, with other levels and modes (e.g., collaborative learning) of education to be further explored as part of future research.

This study and its findings should be interpreted in the light of certain limitations. First, the taxonomy and proposed design patterns and usage implications depend on the literature and data we reviewed. Both qualitative and quantitative empirical data about process mining systems in educational contexts are lacking, and obtaining those data is an overall research need. While much of the present research focuses on the theoretical aspects of process mining in education, few actual field assessments of process mining with deployed systems and users exist (e.g., Mouchel et al. 2023). More specifically, contemporary empirical research looks at process mining systems that may not be tied to real deployment environments (Bogarín et al. 2018; Juhaňák et al. 2019; Cerezo et al. 2020).

In this vein, we are aware that learning data alone may be insufficient to provide a comprehensive overview of learning activities (e.g., Baker and Hawn 2021). We acknowledge that other sources of data, such as tasks and assignments that occur outside of the TML through Moodle, for instance, should be considered to paint a more holistic and complete picture of data in a particular learning context. Also, not all learning activities are captured within digital platforms, which poses a limitation to the completeness and representativeness of the data (Dahlstrom et al. 2014). Data quality issues such as incompleteness, inaccuracy, or inconsistency can severely undermine the insights derived from process mining techniques (Baker and Hawn 2021). In educational contexts, the variability in data entry, the reliance on digital platforms for capturing student interactions, and the absence of standardized data collection protocols across different learning management systems further exacerbate these challenges. Consequently, the effectiveness of process mining in unveiling meaningful patterns and supporting pedagogical decisions may be compromised, necessitating robust data preprocessing and validation methods to ensure reliability and validity of the findings.

In a similar fashion, the comparison with other data sources, such as enrollment or systems data, highlights additional future research avenues (e.g., Dahlstrom et al. 2014). Incorporating such external, contextual data – often referred to as digital traces – can significantly enrich the process mining analysis by providing additional context to student behaviors. The integration of these external data sources requires careful consideration of prerequisites, such as data accessibility and privacy concerns, as well as a thorough understanding of the benefits they can bring to enhance the comprehensiveness of the analysis. Also, our study’s applicability is limited by regional differences and variations in digital maturity among educational providers. The digital transformation of education is unevenly distributed, with significant disparities in the adoption of digital technologies and process mining capabilities across regions and institutions. These variations affect not only the availability and quality of data but also the relevance and applicability of process mining solutions. Consequently, our findings may not be universally applicable, necessitating further research that considers these regional differences and seeks to understand how process mining can be adapted to diverse educational contexts and levels of digital maturity.

While process mining offers unique insights into the flow of learning activities, it is but one of several tools available for analyzing educational data. As mentioned earlier, digital course directories and existing learning systems (e.g., for course registrations or student profiles) exist in practice and offer rich data traces in theory. Such systems provide complementary perspectives, capturing different facets of student engagement and academic progress. The reliance solely on process mining might overlook insights that could be derived from these other sources, suggesting a more integrated approach to data analysis that leverages the strengths of each system to provide a holistic understanding of student learning. Hence, with our study, we aim to establish process mining in education as an additional technique that leverages the experience and research on business process management for educational learning processes. Nevertheless, we believe this should be done in combination with other proven learning analytics methods rather than replacing them. While a large body of reviewed literature focuses on generic learning processes, practical training processes and course-taking sequences provide opportunities for additional empirical investigation, especially when coupled with process mining techniques. An interdisciplinary lens that allows for a multiplicity of data analysis tools offers novel research avenues. These include but are not limited to (1) investigating learner knowledge and skill levels, (2) personalizing systematic learning recommendations for students, and (3) considering learning scenarios in continuous and vocational education, such as MOOCs and practical training.

Exploring these research avenues also requires addressing more fundamental questions. While we have considered the potential of educational process mining as a point of departure, it is important to evaluate its appropriateness compared to other analysis methods and tools. These alternatives might be more suitable to address TML scenarios, for example, in terms of cost-efficiency considerations or questions of data availability. While process mining can offer deep insights into student learning processes, the financial and resource implications of adopting such technologies are not trivial. The costs associated with procuring, customizing, and maintaining sophisticated process mining software, alongside the need for training staff to effectively utilize these tools, can be prohibitive, especially for smaller or resource-constrained educational providers. This limitation underscores the importance of future research on conducting comprehensive cost–benefit analyses to ascertain the feasibility and potential return on investment of process mining initiatives in educational settings. It is necessary to mention and highlight ethical challenges that arise when using digital trace data, particularly in the context of education (Hakimi et al. 2021). As part of our reasoning, an important underlying assumption is that students have consented to the use of their data for TML. Informing and obtaining consent for digital trace data is an extensively discussed ethical challenge. According to Johnson (2019), new personal information might be inferred through already-gathered data. The aggregation and combination of data sources risk de-anonymizing individuals. Similarly, in the context of educational process mining, data might be reused and decontextualized to answer certain questions, thereby potentially developing proxies for certain variables (i.e., gender, race, age) even though those are not explicitly collected. Regulatory considerations also pose a significant limitation to the application of process mining in education. The management and analysis of student data are subject to a complex landscape of privacy laws and regulations, which vary significantly across jurisdictions. Compliance with the General Data Protection Regulation (GDPR) in the European Union, for instance, requires meticulous attention to how student data is collected, processed, and stored. Educational institutions must navigate these regulations carefully, ensuring that process mining activities are conducted in a manner that respects students’ privacy rights and complies with all applicable legal requirements, adding another layer of complexity to the adoption of these technologies. The question of data ownership and consent must be discussed in consideration of the unintended impact educational process mining may have on student learning and social development. Future studies, as well as organizations deploying educational process mining, are required to define privacy notions, such as the extent to which the use of personal data should be accepted by individual learners and the society at large. When turning towards the commercialization of student data, sharing such data with technology vendors also introduces novel issues of inadequate security controls (i.e., Russel et al. 2018). While extant research has explored ethical issues and related solutions for digital trace data in education, research bodies in this realm are fragmented and lack consideration of the education of younger children and informal learning scenarios outside the traditional classroom (Hakimi et al. 2021). This becomes particularly important considering the variety of online learning formats and the increasing availability of various target groups. We urge future research to look further into the ethical challenges of educational process mining and consider the unique affordances of the educational setting when discussing ethical and societal implications.

Table 4 summarizes avenues for future research along our discussions around (1) general questions about the usefulness and use of educational process mining, (2) the analysis of digital learning based on our taxonomy, (3) the use of design patterns for educational process mining, and (4) ethical challenges that can arise in the context of educational process mining.