Multi-Faceted Visual Process Mining and Analytics

Consider the real-time decision-making needed when managing a busy emergency response (ER) department. Patients are coming in and undergo a particular sequence of diagnosis and possibly also treatment steps that can be conceptually captured as a process model, which may change depending on the time of day (working hours vs. after hours) and case load (business as usual vs. state of emergency). The head of the department needs to monitor the current intake, throughput, and related KPIs, like the length of stay (LOS) or ward load (WL) to decide in real time whether to allocate additional resources (activate on-call doctors), to fast-track certain patients (increase their urgency levels), explicitly switch from the usual procedures to the streamlined emergency procedures or back, etc.

To support time-critical decisions, like these in real-time scenarios, we propose to combine streaming process mining with progressive Visual Analytics (VA).

In developing our approach VESPA (Visual Event-Stream Pro[gressive|cess] Analytics), we considered the characteristics of our problem space, which includes SPM, PVA, and expectation of a multi-faceted problem. Our discussion revealed the dimensions of the problem space as provided below. We observed a natural connection between the streaming nature of the process and the value of progressiveness toward providing intermediate (partial) results. The dimensions further allowed us to articulate two research questions relating to timing and appropriateness of the visualization (and interaction).

We outline a user story expressed in two levels of detail to frame and guide our VESPA approach.

• $\mathrm{■}$

  Patient-centric: As an ER administrator, I want to know if the LOS (length of stay) for one ER patient is too high so that I can prioritize them in the waiting queue.

• $\mathrm{■}$

  Ward-centric: As an ER administrator, I want to know if the overall LOS for a cohort of (or all) ER patients is too high or too low so I can adjust the allocation of resources.

This user story manifests in the problem dimensions as below:

1. 1.

   Context: Healthcare

2. 2.

   Task: Primarily we will focus on conformance checking, but this is intended to be augmented with relevant facets

3. 3.

   Data space: A synthetic purpose-built event log has been generated using ChatGPT 4.0 (see Figure 12 and further explanation below)

4. 4.

   Algorithm space: We use behavioral conformance checking (BCC) [7]

5. 5.

   Guards/Rules: Three rules are considered: conformance falling below threshold; a load of an urgent category (critical/high) increasing over a threshold; and the LOS of a given patient increasing over a threshold

6. 6.

   Users: Interchangeable roles of Observers, Searchers, Explorers

7. 7.

   Visualization and Interaction space is aligned with the two user stories of patient-centric and ward-centric with more details provided below

To generate the synthetic dataset ChatGPT has been iteratively queried. The first datasets contains a collection of patient visits to the ER. Specifically, we asked to generate 300 patients according to the following schema: patient ID, condition, urgency level, gender, and age.⁸⁸8The exact prompt for the generation of the patients is: can you create a dataset of 300 patients with the following attributes: Patient id, condition, urgency level, gender, age . conditions include things like flue, headache, fracture, infection, cold, allergy, chest pains, asthma etc. Aim for about 10-12 conditions. Urgency level includes critical, high, medium and low and should make sense for the condition of the patient. Sometimes the same condition can have a different urgency level for example a fracture can be low or critical. Try to generate a variety The distribution of the data resulting from the generation is represented in Figure 13.

Starting from the set of patients, we asked the system to generate sequences of emergency room activities, hinting at the type of activities we are interested in.⁹⁹9The exact prompt for the generation of the activities is: for this patient set, can you generate a log of activities relating to an emergency department. The log contains a timestamp, an activity name and a patient id from the previous dataset. Activities in the beginning can include on route, registration and see a medical professional, activities in the middle can include triage, X-ray, blood test, see a medical professional, and activities at the end can include admission, discharge or leave without being seen. The obtained dataset contained 1,512 events for the 300 patients referring to 268 case variants (so most of the patients followed a unique sequence of activities). On this dataset, some traces were manually removed to avoid particularly meaningless situations and the resulting event log had 770 instances and 142 variants. In the rest of this text, we call this log L1.

Starting from L1, we derived two additional logs, L2 and L3, by applying additional filtering to mimic an off-peak (L2) and an intense scenario (L3). L1, L2, and L3 have also been used to mine the corresponding 3 process models (i.e., M1, M2, and M3) using the approach described in [7]. A picture of the three models built using Fluxicon Disco¹⁰¹⁰10See: https://www.fluxicon.com/disco/. is in Figure 14.

Finally, L1, L2, and L3 have been transformed into three lists of events by sorting each event according to its execution time. With these lists, we constructed our synthetic event stream by concatenating the following lists: L1, L2, L1, and L3 with the intention of simulating a regular daytime period, followed by off-peak (i.e., night), then back to daytime and eventually an intense scenario.

To analyze the data, a streaming process mining pipeline has been implemented using pyBeamline [5]. The pipeline processes each event and computes the following:

• $\mathrm{■}$

  The DFG model [5, 18] updated up to the given point in time;

• $\mathrm{■}$

  The behavioral conformance value [7] of the stream against model M1/M2/M3.

All these values represent the “Process Mining Results Data” in Figure 11, and are combined with the actual raw event, which can be used to compute basic statistics, to form the “Curated Data” (see Figure 11) that is provided as input to the “Visualization and Interaction component”.

The presented VESPA-VIS’s mockups provide an initial illustration of our research questions. That is, when (monitoring or scheduled to more active exploration) and how (views appropriate to patient-centric and ward-centric requirements), progressive visual analytics can best support real-time decision-making in a streaming process mining context. However, there remains a number of further considerations for the proposed approach to be fully realized.

The illustration of the approach through the usecase indicates fertile ground for further developing the approach and robust evaluation to assess the effectiveness of progressive visual analytics for the (real-time) decision support. We anticipate that such an evaluation would require carefully planned user studies with representative participant groups.

Given the continuous nature of the event stream, it is natural to expect a need to “forget” previous event streams when they are no longer relevant for the current decision making. So far, we have considered a rather straightforward way of simply ‘forgetting’ patients who have exited the ER through discharge or transferal to another ward. Yet this prevents the head of the ER department from comparing the currently observed situation with previously observed situations – for example, the processes occurring on a current New Years holiday day to the processes on the same day in previous years – to identify best practices or simply “what has worked in the past”. Identification of ‘forgetfulness’ thresholds is in itself a complex and multi-faceted problem that requires further work, although prior literature gives hints (see for example [15]).

Although the focus of the approach is to support real-time decision making, the insights gained from the proposed approach present an opportunity to inform process enhancement. Exploring this opportunity requires further consideration.

Like interactive visual data analysis in general [19], process mining in particular is exploratory and human-driven. Process analysts typically have to engage in iterative exploration cycles of process visualizations (see Figure 17) to build an understanding of the process, examine different scenarios, and generate hypotheses [26]. Hypotheses are then tested, refined, or discarded based on intermediate insights, which, in turn, guide analysts in choosing what to explore next [18] and lead to the crystallization of knowledge [21].

Many process mining tools support this interactive exploration through components such as filter masks or sliders that allow analysts to create views and isolate data subsets of interest. However, these interactions often lack transparency and context, making it difficult for users to anticipate the effects of their actions before executing them (What should I usefully do?) and understand the effect once an action has been executed (Have I achieved the desired outcome?).

As an example, consider a process mining analyst, let’s name him Bob, who is in charge of analyzing a Road Fine Management process [7] visualized as a directly-follows graphs (DFG). His goal is to investigate cases where offenders do not pay their fines. Bob’s analysis steps are sketched in Figure 18.

As a first step, Bob loads the raw data provided by the police ($L_0$) into a process mining tool. His goal is to get an initial understanding of the structure of the process. To achieve this goal, he intents to focus on the most common behavior using a variant filter to remove infrequent cases. However, to find a suitable abstraction level for his analysis, Bob has to go through the costly procedure of applying the filter three times ($o_1,o_2,o_3$). He selects different filtering thresholds of 75%, 85%, and 90% of the cases in the log, and each time he inspects the resulting DFG ($v_1,v_2,v_3$) to assess visually the effects of the filters. While the first two filter configurations remove too many cases, he settles for $o_3$.

This interactive selection of most frequent cases is a common first step of many process mining analyses. However, during this process, Bob runs into several limitations:

• $\mathrm{■}$

  There is no preview mechanism to support the decision for a suitable filter threshold: Bob must fully apply each filter to see what the result looks like;

• $\mathrm{■}$

  Comparisons across multiple visualizations resulting from the filtering are lost: Since each filter operation completely replaces the DFG with a new one, Bob is forced to take screenshots or compare from short-term memory;

• $\mathrm{■}$

  The DFG shows the impact of his filter on the control-flow only: Bob would need to create different views of the data to see how his filtering impacts other facets of the process.

The difficulties Bob faces demonstrate some of the challenges that stem from limitations of interaction functionalities currently being used in PM practice. These limitations introduce constraints to the process of process mining (PPM) [18] and potentially hinder performance when making sense of event log data. The objective of our working group is to investigate these challenges and outline the opportunity of enhancing the process exploration based on established concepts, methods, and techniques from the realm of VA.

With our working group, we aim to improve the overall process of making sense during the PPM. To this end, we compiled relevant previous work from the PM and VA communities, both to understand some of the cognitive challenges during process exploration and to bring together models and approaches that can inform the design of advanced process exploration techniques to overcome these challenges.

During the group discussion, we considered several works from PM and VA. These works are concerned with the cognitive, processing, and interactive mechanisms that are relevant during visual process exploration.

In an attempt to pinpoint the fundamental conceptual aspects of the desired process exploration support, we came up with the following (incomplete) list of notations inspired by the section on interactive selection and accentuation in [19]:

• $\mathrm{■}$

  $D$, the data to be visualized, explored, and understood

• $\mathrm{■}$

  $D_{+}\subseteq D$, the currently relevant focus data, subject to change frequently during process exploration

• $\mathrm{■}$

  $D_{-}=D\setminus D_{+}$, the data currently not being of relevance for the process exploration

• $\mathrm{■}$

  $S$, a state capturing the data underlying the visualization

• $\mathrm{■}$

  $S_{cur}$, the “current” state

• $\mathrm{■}$

  $S_{old}$, the “old” state

• $\mathrm{■}$

  $\{S_{a_1},S_{a_2},\mathrm{\dots },S_{a_n}\}$, a set of possible (useful) “alternative” states that can be entered through alternative interactions

• $\mathrm{■}$

  $\delta (S_i,S_j)$, the explicit difference(s) between two states

• $\mathrm{■}$

  …

Based on these notations, we defined exploration as the repeated refinement and change of $D^{+}$ (and $D_{-}$ respectively), which usually involves numerous state changes (e.g., the three different filtering states in Bob’s exploration example). Moreover, it seems that understanding state changes is crucial for effective exploration. Possible options for supporting the understanding of state changes can be based on Gleicher et al.’s [10] strategies for visual comparison:

• $\mathrm{■}$

  Juxtaposition: Visualize $S_i$ and $S_j$ side by side

• $\mathrm{■}$

  Superposition: Superimpose the visualization of $S_i$ over the visualization of $S_j$

• $\mathrm{■}$

  Explicit encodingg: Visualize $\delta (S_i,S_j)$ directly

So far, these strategies are not sufficiently integrated into existing process exploration practice!

To understand the users’ needs better, we further conceptualized a cycle of interaction for a seamless analysis that is based on feedback and feedforward. Generally, in interactive exploration/analysis, the analyst interacts with the visualizations for dicing, slicing and relating different parts of the data. So, the exploration cycle starts with the analyst expressing an intent to change the visual representation. Yi et al. [24] identified several different categories of interaction intents called “Show me…”, of which we focus on the intents related to changing the focused subset of the data exploration:

• $\mathrm{■}$

  Show me something else: A different subset of the data (e.g., navigate in time) will be visualized; $D_{+}\to D_{+}^{\prime }$.

• $\mathrm{■}$

  Show me more/less: A subset of different size or level of aggregation (e.g., reduce number of nodes in DFG) will be visualized; $|D_{+}|\ne |D_{+}^{\prime }|$.

• $\mathrm{■}$

  Show me something conditionally: A subset that fulfills certain (filter) condition(s) (e.g., filter for frequent variants) will be visualized; $C(D_{+})=true$.

• $\mathrm{■}$

  Show me related things: A subset that is (in some way) related to the currently shown subset (e.g., brushing and linking across multiple faceted views) will be visualized; $R(D_{+},D_{+}^{\prime })$, typically $D_{+}\sim D_{+}^{\prime }$.

These intents lead to interactions to which the system responds by providing visual feedback, and what we would like to emphasize, also visual feedforward. We envisioned the following scenario in which a user executes an interaction and receives the corresponding visual feedback and feedforward.

1. 1.

   The user’s intention is typically communicated to the system through different ways of interaction (e.g., hovering a visual mark in the visualization or clicking a button or slider in the user interface).

2. 2.

   The system interprets the user’s action and then provides relevant context and suggests possible next steps with previews of their impact. Here we can explore a large design space of different useful visual feedback and feedforward, which generally are dependent on the semantics of the interaction and will incorporate different facets of the data. The additional context and possible next steps help the analyst to decide if the intended interaction outcome has been obtained, and if not, how to execute their alternative more fruitful interactions.

3. 3.

   Based on the visual feedback and feedforward, the analyst can now better understand interaction effects and can more easily decide what to do next. The cycle starts again as the analyst continues the exploration and expresses their new intents through new interactions.

With this general scenario now being clear, the question that remains is how to design the visual feedforward and feedback concretely. However, given the huge design space, this is quite a challenging task.

For our design sketches, we drew inspiration from previous work on enhancing interaction with visual feedback and feedforward. In particular, we considered:

• $\mathrm{■}$

  Scented widgets [23, 5] embed small miniature visualizations (e.g., histograms) directly into graphical control elements such as sliders.

• $\mathrm{■}$

  Small multiples and large singles [20] is a concept to preview thumbnails of alternative parameterizations of visual representations.

• $\mathrm{■}$

  Guidance visual cues [9] can be embedded into visualization views to indicate potentially interesting next navigation targets.

• $\mathrm{■}$

  Octopocus [1] is an interaction technique that provides feedforward as an interactive gesture is performed to indicate possible interaction outcomes.

The sketches used these inspirational techniques to outline possible solutions for informative visual feedback and feedforward for process exploration. Figure 20 shows a selection of our sketches, including conceptualization of interaction intents, general interface ideas, scented slider widgets, and preview thumbnails. From these and further similar sketches, we abstracted the following design dimensions that can play a role when implementing enhanced process exploration mechanisms:

Interaction control:

What type of control is used to carry out the interaction? Slider, button, hover area, spoken command, etc.

Interaction integration:

Where is the interaction control located? Integrated in the visualization vs. external to the visualization in a separate user interface.

Visual feedback/feedforward integration:

Where is the visual feedback/feedforward shown? Visualization enhancement integrated into the visualization vs. Interface enhancement integrated into the interaction control.

Process exploration constitutes a fundamental task in process mining, primarily aimed at facilitating the exploration and generation of hypotheses about the underlying process behavior captured in event data. In particular, during the initial phases of process mining projects, analysts engage in various exploratory activities – they dedicate time to familiarize themselves with the data, develop a preliminary understanding of the process, formulate or refine analytical questions, and uncover unexpected patterns or insights [6]. This iterative exploration plays a crucial role in shaping the direction of subsequent hypothesis testing [5].

The way process exploration is currently performed typically involves the use of discovered Process Maps, such as the Directly-Follows Graphs (DFGs) [27, 9, 17], where nodes represent activities and edges indicate direct successions based on event log timestamps. Using the DFG as a visual representation to enable exploration generally implies that the primary facet (i.e., the first-class citizen of our analytical workflow) is the temporal order of activities [16].

More detailed insights are often introduced through additional visual channels, such as projecting performance metrics (e.g., activity duration, waiting time) or frequency information (e.g., how often transitions occur) on top of the DFG structure. These additional visual cues enrich the visualization but do not alter the primary facet. This allows the analyst to gain insight on different aspects of the process, such as identifying activities (e.g., activities with high duration, infrequent activities, endpoint activities), fragments (e.g., the most frequent fragment), transitions (e.g., transitions with high durations), or bottlenecks [10].

Less commonly, the DFG can be restructured entirely by using resources as the primary facet [11]. In resource-centric DFGs, nodes represent individuals, roles, or organizational units, and the edges reflect handovers or collaborative interactions. This shift enables analysis of organizational dynamics, revealing patterns such as teamwork structures, silos, or handover inefficiencies.

Despite the value of these perspectives, a major limitation is that the primary facet is typically fixed, constraining the flexibility of exploration. Real-world analytical tasks often require users to continuously shift between facets [17] – from case-centric to activity-centric, from control flow to resource flow, or from process variants to attribute analysis in order to discover inter- and intra-dependencies. Fixing the facet can obscure relevant patterns, reduce user agency, and lead to a cognitive mismatch when visualizations do not align with the user’s current task or mental model [13].

To truly support hypothesis generation and sense-making, process exploration tools must support flexible transitions between facets, allowing analysts to dynamically reframe the data perspective depending on their line of inquiry [17]. Typically, users are interested in exploring similarities and differences between (combinations of) facets, i.e., the visual analysis approach should support more complex comparison tasks. To support this, dimensionality reduction might help here; it is flexible in what facets are considered and helps to reveal similarities (local neighborhoods) in the high-dimensional space.

We investigated methods for analyzing event logs from multiple, complementary facets [2], with the goal of decomposing them into smaller, more interpretable subsets. These event log subsets can then be explored in greater detail using compact and more readable representations such as DFGs. To this end, we experimented with both topic modeling [12] and embedding-based techniques [1]¹¹¹¹11https://va-embeddings-browser.ivis.itn.liu.se, applying various strategies to represent event logs either as textual documents (for topic modeling) or as structured feature vectors (for embedding). In particular, we started exploring representations that capture both the composition of logs as collections of activities and as sequences of transitions between activities. These representations have potential to account not only for the temporal order in which activities occur, but also incorporate quantitative time aspects such as activity durations and additional contextual attributes. Finally, we explored the possibilities that multiple coordinated views can provide as a way to enhance the analysis process.

The result of these investigations is summarized in the workflow presented in Figure 22. In the following, we describe each of the steps involved in the workflow.

As part of future work, we plan to implement a comprehensive visual analytics system that builds on our current approach. In the current system design, each dot in the visualization represents a case positioned using DR techniques. Moving forward, we aim to explore alternative semantic representations for dots, such as activities or process variants. Moreover, we would like to systematically investigate how to encode activities, variants, and traces in a multi-faceted manner to best support various exploratory goals. To further enrich the visual expressiveness, we plan to integrate glyph-based representations that convey additional attributes or contextual cues (e.g., for selected subpopulations). Moreover, we intend to integrate our approach with existing process discovery algorithms to support the creation of DFGs for selected subpopulations. Additionally, we intend to expand the visual space to support additional facets, including relationships, control-flow, and resources. Another interesting research direction is to study ensemble methods for embeddings where various embedding approaches might be combined (either conceptually different state-of-the-art embedding technologies or the same embedding algorithm with various hyperparameter settings) to provide better performance [15]. A key objective of our future work will be to evaluate the effectiveness of the proposed system in supporting exploratory process mining, demonstrating its ability to accommodate a wide range of analytical tasks and user needs.

The surge in the digitalization of processes has catered for an abundance of recorded data. Process mining successfully leverages this renownedly precious information source for knowledge acquisition and enhancement, thus garnering significant attention in research and industry alike. Digitalization, on the other side, has also spurred the transcend of business landscapes beyond the silos of single processes, departments, or organisations [1]. While favouring the hyperconnection of expertise and knowledge, it has also exposed the limits of classical process mining, whose techniques are typically designed to focus on those silos. Events like the outbreak of pandemics in healthcare, the domino effect of stock market agitations in finance, and the blockades caused by issues along logistic routes in supply chain management, disruptively illustrated the expansion in space and time of events well beyond the boundaries of their locality.

We are experiencing a more and more predominant passage from an architecture of processes [2] to what we call a process ecosystem. Recent advances in the scientific literature call for multi-perspective, inter-instance, cross-process approaches [3]. Beyond the technical objectives of efficacy and scalability for this new wave of approaches, we claim in this paper that an expressive, appropriate, effective analysis of data from hyperconnected settings is key to pursuing the ultimate aim of knowledge extraction. To this end, we advocate the integration of process analytics to propel the potential of process mining and significantly contribute to its capability to handle process ecosystems.

To motivate our work, we consider the organization of a scientific conference. To achieve this, the process of reviewing papers has to be enacted. The process is organized as follows: authors create new submissions by sending their abstract and full paper. The program chair then assigns the papers to review to program committee (PC) members. PC members then review the papers, potentially delegating the task to subreviewers. Once the review deadline expires, PC members begin discussing their reviews with senior PC members. Once the discussion is over, program chairs take the final decision on the paper and, in case of acceptance, the authors submit a camera-ready version of their paper.

This process encompasses several interrelated activities, such as submitting papers, assigning papers, writing reviews, discussing reviews, and taking decisions. Such activities involve different resources, such as papers, reviews, and comments. They are also performed by actors under different roles, such as authors, (senior) PC members, and program chairs. It is also worth noting that the same resource may participate in different activities carried out by different actors, and the same actor may perform different activities, or the same activity multiple times, depending on the involved resource.

To support the review process, the conference organizers rely on a Conference Management System (CMS). Instead, other supporting processes, such as participating in PC meetings and discussions, happen outside the CMS through other channels, such as email or videoconferencing applications. These processes are influenced by and influence the review process, as they all share a subset of actors and resources. Figure 27 shows an Enterprise Architecture [4] model that captures the activities, the objects, the supporting applications, and the roles that the actors have in this example.

• $\mathrm{■}$

  Multiplicity of perspectives: Unlike analysis involving a single stakeholder or a particular part of a process, process ecosystems involve processes representing multiple perspectives within a complex system. Analyzing these multiple perspectives require creating multiple representations that enable switching from one perspective to the other seamlessly and with methods allowing cross-perspective analysis.

• $\mathrm{■}$

  Inter-process interactions: An inherent challenge with multiple processes in process ecosystems is the diverse range of interactions that can take place within the processes. For instance, processes can be interdependent on one another, could be competing for shared resources or processes can be sub-parts of larger processes. These range of relations require an analyst or a designer to devise bespoke computations or visualizations that can handle a large range of complex relations between processes.

• $\mathrm{■}$

  Temporal constraints: Temporal constraints and rules do already require special attention within analysis involving a single process. With multiple processes interacting in diverse ways, the effort to oversee temporal constraints is exacerbated. Analysts require multi-faceted but temporally-aligned representations to be able to concurrently analyze multiple interacting processes.

• $\mathrm{■}$

  Data heterogeneity and quality issues: With several processes that needs to be interlinked and analyzed concurrently, the data that captures these process will be diverse and be available in incompatible formats, requiring extensive effort in data processing. With data being captured in multiple levels, the probability of data errors and gaps occurring is also higher.

• $\mathrm{■}$

  Dynamic evolution: One inherent character of complex inter-related systems is their continuous evolution. This could mean that dependencies across entities could change, new dependencies could emerge and resource needs and pathways of processes could alter. This requires dynamic representations that enables the monitoring and analysis of evolving systems.

Effective monitoring and analysis of process ecosystems requires a combination of techniques that can extract structured knowledge from event data and support interactive, human-centered exploration. Process Mining (PM) and Visual Analytics (VA) offer complementary strengths that, when combined, enable a powerful approach to understanding such complex systems.

• $\mathrm{■}$

  Linking Process Discovery with Interactive Exploration: PM excels at deriving structured process models from event data, revealing control-flow patterns and performance metrics. VA complements this by enabling users to interactively explore these models and the underlying data, focusing on aspects most relevant to their questions.

• $\mathrm{■}$

  Integrating Temporal Representations and Analysis: PM tools provide time-based metrics and can segment process stages temporally. VA enriches this with expressive visual metaphors, such as timelines, storylines, Marey charts, or dynamic graphs, that allow users to perceive the progression, synchronicity, and timing of multiple processes in context.

• $\mathrm{■}$

  Revealing Inter-Process Interactions: While PM can identify shared activities or resources, VA supports the visualization of dependencies, influences, and synchronizations across process instances. This helps users reason about how processes interact, support, compete with, or block each other.

• $\mathrm{■}$

  Supporting Multi-Actor and Multi-Perspective Analysis: In process ecosystems involving many stakeholders (e.g., coordinators, reviewers, chairs), PM can identify roles and responsibilities. VA allows for flexible filtering and perspective switching, helping stakeholders understand the system from their own or others’ viewpoints.

• $\mathrm{■}$

  Facilitating Monitoring and Timely Intervention: The combination enables the construction of rich monitoring dashboards where PM provides the structured event flow and derived KPIs, while VA supports intuitive visual layouts and alerts that guide attention to issues such as delays, coordination problems, or resource bottlenecks (see Figure 29).

• $\mathrm{■}$

  Managing Complexity and Enhancing Interpretability: PM’s algorithmic capabilities scale to large datasets, and VA provides ways to abstract, aggregate, and interactively refine views, allowing users to navigate complexity without losing important details.

• $\mathrm{■}$

  Enabling Retrospective Learning and Knowledge Externalization: PM can identify patterns of behavior over historical data, and VA supports sensemaking and storytelling helping analysts communicate findings, compare scenarios, and build shared understanding of systemic dynamics.

In this work, we highlighted the need to move beyond traditional, isolated views of processes and embrace the concept of process ecosystems – systems of multiple interrelated and interacting processes evolving in parallel. We argued that understanding such ecosystems requires not only the discovery of individual process models but also tools for monitoring, visualizing, and analyzing their temporal development, mutual dependencies, and interactions.

We discussed how the synergy between Process Mining (PM) and Visual Analytics (VA) can address this need. PM offers robust methods for discovering and quantifying processes based on event data, while VA brings in powerful techniques for time-oriented representation, interactive exploration, and multiscale analysis. Time-based visualizations, coordinated multiple views, and interaction techniques such as brushing, filtering, and dynamic level-of-detail provide essential support for making complex process interrelations observable and interpretable.

Using the example of scientific conference organization, we illustrated a process ecosystem and discussed some key goals an visual process analysis approach could support. We also pointed to scalability challenges and the importance of grouping and abstracting processes using clustering based on task-specific similarity metrics.

Future research directions include:

• $\mathrm{■}$

  Developing scalable visual representations that combine aggregate views with detailed process trajectories.

• $\mathrm{■}$

  Designing similarity measures and clustering techniques tailored to different types of process interactions (e.g., synchronization, resource sharing, interference).

• $\mathrm{■}$

  Supporting dynamic grouping and focus+context views to enable exploration of evolving sub-ecosystems.

• $\mathrm{■}$

  Integrating causal inference techniques to understand not only correlations but also influence among processes.

• $\mathrm{■}$

  Creating domain-specific dashboards that bring together PM and VA components to support monitoring and decision making in real-time.

• $\mathrm{■}$

  Conducting empirical studies with domain experts to evaluate usability and effectiveness of the proposed approaches.

By advancing in these directions, we can build more intelligent and adaptive tools for managing complex systems of interacting processes across domains such as healthcare, logistics, manufacturing, and scientific workflows.