# Gyunam Park > Gyunam Park, assistant professor at Eindhoven University of Technology. Process mining at the intersection of data science and process science: object-centric process mining, predictive process monitoring, process simulation and neuro-symbolic AI, with applications in manufacturing and healthcare. Tenured Assistant Professor at TU/e | Group Lead at Fraunhofer FIT. This is the research site of Gyunam Park (https://www.gyunam.com/). Every page is also served as Markdown at its own URL plus `index.md`. The whole site in one file is https://www.gyunam.com/llms-full.txt. Structured exports (JSON, BibTeX) are under https://www.gyunam.com/data/. --- # Gyunam Park - Type: Profile - URL: https://www.gyunam.com/ - Role: Tenured Assistant Professor at TU/e | Group Lead at Fraunhofer FIT I am an assistant professor in the [Department of Mathematics and Computer Science](https://www.tue.nl/en/our-university/departments/mathematics-and-computer-science) at Eindhoven University of Technology, where I lead a research line on neuro-symbolic AI in [Process Analytics](https://pa.win.tue.nl/) group. I also lead a process mining group at [Fraunhofer FIT](https://www.fit.fraunhofer.de/en/business-areas/data-science-and-artificial-intelligence/process-mining.html). I received my PhD in Computer Science from RWTH Aachen University with the thesis *Action-oriented Process Mining*, my MSc in Industrial and Management Engineering from POSTECH, and my BSc in Technology Management and Computer Science from UNIST. ## Research vision Organizations record digital traces of who did what, when, and for which case. My research develops data science methods that learn from these traces to predict and steer what happens next, in a way that organizations can trust. The methodological core is **neuro-symbolic process learning**: learning systems that absorb explicit process knowledge, such as clinical guidelines, safety regulations and documented procedures, both as a prior that shapes what a model believes and as a guardrail that bounds what it predicts and recommends. The foundation is process science, machine learning and statistics. The application focus is healthcare, manufacturing and public-sector organizations, where unconstrained learning is unusable and pure rule-following is too rigid. ## Research lines - [Neuro-symbolic process learning](/research-line/neuro-symbolic-process-learning/): the methodological core. Formal process knowledge from guidelines, regulations and procedures enters learning systems as a prior and as a guardrail, for prediction, recommendation and anomaly detection. - [Object-centric process mining](/research-line/object-centric-process-mining/): extraction, standards, analysis and tooling for event data in which several objects interact, from SAP ERP to manufacturing and healthcare. - [Prescriptive process analytics and digital twins](/research-line/prescriptive-process-analytics/): predictive monitoring, recommended actions, simulation and digital twins of organizations that close the loop from insight to improvement. - [Process discovery and diagnostics](/research-line/process-discovery-and-diagnostics/): sound yet expressive process models, decomposition, performance patterns, deviations and resource effects, the descriptive side that the other lines build on. ## Off the screen As much as I enjoy doing research sitting in front of a monitor, I enjoy running, cycling and swimming just as much. I like listening to music, even better with people and drinks. I am proud of the kimchi I make with Marlene, though it will never compare to my mom's. ## Honors and awards - 2026: Best Paper Award, BPM conference (Toronto, Canada) - 2024: Best Dissertation Award Runner-Up, BPM conference (Seville, Spain) - 2023: Best Paper Award, EduPM Workshop (Rome, Italy) - 2022: Best Paper Award, PQMI Workshop (Bolzano, Italy) - 2011–2017: 8 Consecutive Scholarships, Academic Performance Scholarship (S. Korea) - 2015–2016: Scholarship, Samsung Dream Class Scholarships (S. Korea) - 2013: 3rd-prize Award, Global Startup Competition (S. Korea) --- ## Research lines --- # Neuro-symbolic process learning Injecting formal process knowledge from guidelines, regulations and documented procedures into learning systems, as a prior that shapes what a model believes and as a guardrail that bounds what it predicts and recommends. - Type: Research line - URL: https://www.gyunam.com/research-line/neuro-symbolic-process-learning/ ## Practical problem Healthcare, manufacturing and public-sector organizations hold explicit process knowledge: clinical guidelines, safety regulations, documented procedures. They also record event data about what actually happens. Today the two are used separately. Models trained on the event data alone are unusable in these settings, because they can predict and recommend behaviour that violates rules the organization must follow. Pure rule-following is too rigid, because the rules live in text and in people's heads, some are outdated or contradicted by practice, and they say nothing about what is likely to happen next. Since 2024 I have built a research line on injecting formal process knowledge into learning, so that models absorb that knowledge both as a prior and as a guardrail. ## Research questions The line has two pillars. **Process knowledge engineering.** Process knowledge exists, but it has to be acquired, trusted and maintained. - How can rules be acquired from documents, such as guidelines and regulations, and from event data, and how should each rule be scored on data support, consistency, currency and fidelity to its source? - When a rule and the data disagree, is the rule stale or is the behaviour anomalous? - How are rules versioned under concept drift, including rules distilled back from trained models? Rule-quality estimation, conflict resolution and drift detection are statistical inference problems, so the statistical questions of the line concentrate here. **Neuro-symbolic process analytics.** The central open problem is making neural models respect temporal rules about ordering, timing and deadlines, beyond the rules about what happens that current methods support. - How should formal knowledge enter predictive, prescriptive and unsupervised methods: as a prior that shapes the model, as a guardrail that bounds its outputs, or both? - Which benchmark of paired datasets, where event data and formal knowledge are available together, lets the two mechanisms be compared empirically? - For complex systems with many interacting objects and resources, how can knowledge-informed graph neural networks learn on top of a knowledge graph of explicit process structure? ## Contributions The contributions so far fall into six groups, each linked to the publication in which it appears. **Acquiring process knowledge from text and from data.** - A pipeline that translates each sentence of a guideline or process description into a Declare constraint with a fine-tuned open large language model and merges the constraints into a Declare model; fine-tuning beats both prompting and the rule-based state of the art ([Decision Support Systems, 2026](/publication/2026-dss-declarativenl/)). - An automated knowledge pipeline that builds a SNOMED CT knowledge graph from the Surviving Sepsis Campaign guidelines, mines Horn rules from it and translates them into first-order logic axioms ([Engineering Applications of Artificial Intelligence, 2026](/publication/2026-eaai-sepsis/)). - A scheme that mines control-flow, temporal and payload rules from event data with Declare mining, service-level-agreement analysis and statistical correlations, formalizes them in Linear Temporal Logic and first-order logic, and categorizes them into class-dependent and class-independent knowledge ([PAKDD 2026](/publication/2026-pakdd-twostage/), [CAiSE 2026](/publication/2026-caise-complianceppm/)). - Rule selection: a gating score from the mean and variance of rule satisfaction that prunes harmful rules after pretraining ([PAKDD 2026](/publication/2026-pakdd-twostage/)), and a significance test that keeps only mined temporal patterns whose satisfied and violated cases differ on an outcome score ([CAiSE Forum 2024](/publication/2024-caise-f/)); earlier groundwork extracted rules for study planning from event data ([ICPM Workshops 2023](/publication/2023-icpm-edupm/)). **Knowledge-informed prediction.** - A four-stage neuro-symbolic pipeline for predictive process monitoring that injects rules through Logic Tensor Networks, with three injection mechanisms relative to the classification predicate: feature expansion at the input, output refinement and parallel constraints as auxiliary objectives ([CAiSE 2026](/publication/2026-caise-complianceppm/)). - Two-stage training that weights data axioms above knowledge axioms during pretraining, prunes rules and fine-tunes on the pruned knowledge base, which avoids the collapse that unguarded logic losses cause ([PAKDD 2026](/publication/2026-pakdd-twostage/)). - NeSy-SMP, a real-time sepsis mortality predictor that grounds Logic Tensor Network predicates in a bidirectional LSTM over clinical event sequences, with a dual axiom architecture, weak anchoring of clinical predicates to established thresholds and a three-tier explainability framework ([Engineering Applications of Artificial Intelligence, 2026](/publication/2026-eaai-sepsis/)). **Knowledge-informed recommendation and planning.** - NeSyPA, a prescriptive method that combines a Transformer prescriptor with Declare constraints compiled into automata, relaxes soft constraints hierarchically only when no activity is feasible, and distils feasibility-masked, outcome-weighted targets into the model ([BPM 2026, Best Paper Award](/publication/2026-bpm-nesyprescriptive/)). - Behavioural recommendations mined from event logs and fed into AI planning as soft constraints next to hard regulatory rules, with optimal plans found by integer linear programming ([CAiSE Forum 2024](/publication/2024-caise-f/)). **Knowledge-informed anomaly detection.** - A denoising autoencoder fine-tuned with Logic Tensor Networks so that reconstruction loss is combined with the satisfaction of mined Declare constraints, which improves detection with as few as ten rare but conformant traces ([CAiSE 2026](/publication/2026-caise-nesyanomaly/)). **Benchmarks, metrics and evaluation protocols.** - A public benchmark of 969 sentences paired with expert-annotated Declare constraints across healthcare, compliance, education and finance, and an error metric that weights template and direction errors above activity errors ([Decision Support Systems, 2026](/publication/2026-dss-declarativenl/)). - Compliance-aware evaluation protocols that enrich test sets with rule-compliant traces and measure constraint satisfaction alongside accuracy ([CAiSE 2026](/publication/2026-caise-complianceppm/), [PAKDD 2026](/publication/2026-pakdd-twostage/)). - The Effective Compliance Rate, which counts recommendations that are both outcome-improving and constraint-compliant ([BPM 2026](/publication/2026-bpm-nesyprescriptive/)). - Clinical operating-point analysis and external validation on a second hospital dataset for the sepsis predictor ([Engineering Applications of Artificial Intelligence, 2026](/publication/2026-eaai-sepsis/)). **Synthesis.** - A tutorial that organizes neuro-symbolic work in BPM along Kautz's architectural taxonomy, analyses failure modes such as vacuous satisfaction and the soft-compliance gap, and gives a five-question decision aid for when the methods pay off ([BPM 2026](/publication/2026-bpm-nesybpm/)). --- # Object-centric process mining Process mining for processes in which orders, items, invoices, machines and patients interact, instead of one isolated case at a time. - Type: Research line - URL: https://www.gyunam.com/research-line/object-centric-process-mining/ ## Practical problem Classical process mining assumes that every event belongs to exactly one case. Processes run on ERP systems, manufacturing lines and hospital information systems break this assumption: an order has many items, deliveries and invoices, a technician's day spans many service schedules, and a single event often concerns several of these objects at once. Forcing such data into one case notion creates convergence and divergence errors and hides exactly the interactions where delays, overstock and compliance problems arise. Object-centric event data removes this blind spot, but it needs its own standard, extraction methods, analysis techniques and tools. Since 2020, I have built this chain from the raw tables of SAP ERP to object-centric event logs, analysis techniques and open-source software. ## Research questions The line has three pillars. **Standards and extraction.** Object-centric event data has to be stored, exchanged and obtained before it can be analysed. - How should event data in which one event touches several objects of different types be represented and exchanged without losing the object relationships, and how can it be stored and queried at the scale of millions of events? - How can object-centric event logs be extracted from the thousands of tables of an ERP system such as SAP, or from a clinical data model such as OMOP, with as little domain knowledge as possible, and how does an analyst find out which processes a database supports at all? **Analysis techniques.** The notions that process mining relies on change when cases overlap. - Which definitions of performance, variant, conformance and constraint violation make sense for interacting objects, and how are they computed by replaying an object-centric Petri net or an object-centric directly-follows graph? - How can object-centric event data feed machine learning while keeping the graph structure of a process execution, instead of flattening it into a table or a sequence? **Practice.** Object-centric analysis has to pay off in real projects. - How does a process mining project methodology such as PM² change for object-centric projects, and what do real purchase-to-pay, after-sales and healthcare processes reveal when analysed this way? ## Contributions The contributions so far fall into five groups, each linked to the publication in which it appears. **The OCEL standard and its storage.** - The OCEL standard for object-centric event logs: a meta-model in which each event references a set of objects of different types and object attributes are stored once, with JSON and XML serializations, validation schemas and Python and Java tool support ([ADBIS Workshops 2021](/publication/2021-simpda/)). - A MongoDB schema and indexing scheme for OCEL with database-level aggregations for object lifecycles, statistics and object-centric directly-follows graphs, assessed on synthetic logs of 1 to 100 million events ([ICPM Demo 2021](/publication/2021-icpm-demo/)). **Extraction from SAP ERP and clinical databases.** - A technology-independent approach for extracting object-centric event data from databases: a graph of relationships derived from the data dictionary, blueprints that map table entries to events, a formal conversion to OCEL and the open-source Interactive SAP Explorer with two published order-to-cash and purchase-to-pay logs ([Journal of Intelligent Information Systems, 2023](/publication/2023-jiis-generic-sap-extraction/)); first presented as an extraction approach ([ICPM Workshops 2021](/publication/2021-pqmi/)) and as an interactive process identification demo ([ICPM Demo 2022](/publication/2022-icpm-demo/)). - OMOP CDM4PM, a formally defined subset of the OMOP common data model, and a transformation into OCEL whose event and object types are derived from the OMOP concept table and the SNOMED CT hierarchy, implemented in omop2ocel and evaluated on three MIMIC-IV sub-datasets ([Journal of Biomedical Informatics, 2024](/publication/2024-jbi-omop/)). **Performance, constraints and conformance on object-centric data.** - OPerA: replay of a discovered object-centric Petri net to compute flow, sojourn, waiting and service time per event occurrence, plus the object-centric synchronization, pooling and lagging time ([ER 2022](/publication/2022-er/)). - Object-centric constraint graphs and behavioral metrics for monitoring control-flow, object involvement and performance constraints, with the ProPPa web application ([ICPM Workshops 2022, Best Paper Award](/publication/2022-pqmi/)), and constraint checking for object-centric process executions ([ICPM Workshops 2023](/publication/2023-icpm-pqmi/)). - A taxonomy of object-centric operational problems, Operational Problem Graphs with formal occurrence semantics and a monitoring function, and the ProAct monitoring engine ([Computers in Industry, 2025](/publication/2024-comind/)). - Conformance checking and performance analysis on object-centric directly-follows graphs, which project per object type, align the projected log and map log and model moves back onto the graph, in the Explori tool ([BPM Forum 2024](/publication/2024-bpm-forum/)). - Super variants, which summarize object-centric variants with choice, optionality and object multiplicity through a hitting-set selection and a matching-based merging hierarchy ([BPM 2024](/publication/2024-bpm/)). - GOProQ, a graphical query language with formal syntax and semantics over object-centric process executions and a query-reordering evaluation strategy, evaluated on six public OCELs ([ICSOC 2025](/publication/2025-icsoc-goproq/)). **Object-centric data for machine learning.** - A framework that translates control-flow, data, resource and performance features to object-centric event logs, adds an object perspective, and defines tabular, sequential and graph encodings of process executions ([ICSOC 2022](/publication/2022-icsoc/)). - An approach that keeps process executions as event graphs throughout the machine-learning pipeline, as graph neural network input or as graph embeddings, with an evaluation protocol that isolates the value of structure alone ([Engineering Applications of Artificial Intelligence, 2023](/publication/2023-eaai/)), and object-centric process mining to improve predictive process monitoring ([ECIS 2024](/publication/2024-ecis/)). **Tools and case studies.** - ocpa, the open-source Python library for object-centric process analysis: OCEL import and export, process executions and variants, object-centric Petri net discovery, precision and fitness, constraint monitoring, performance analysis and feature extraction ([Software Impacts, 2022](/publication/2023-softwareimpact/)). - An extension of the PM² project methodology to five stages for object-centric projects, applied to the after-sales service process of Borusan Cat ([ICPM Workshops 2023](/publication/2023-icpm-pqmi-2/)) and to a real-life purchase-to-pay process ([International Journal of Data Science and Analytics, 2025](/publication/2025-ijdsa-p2p/)). - Overstock in the purchase-to-pay process of Europe's largest pet retailer: a conceptual model with five object types, extraction from Celonis into OCEL, two discovered object-centric Petri nets and two improvement measures agreed with logistics management ([CAiSE Workshops 2024](/publication/2024-caise-w/)). --- # Prescriptive process analytics and digital twins From predicting process problems to recommending actions and testing them in simulation models and digital twins of organizations before they reach the real process. - Type: Research line - URL: https://www.gyunam.com/research-line/prescriptive-process-analytics/ ## Practical problem Insights do not improve a process; actions do. Most process mining stops at dashboards and leaves the translation into decisions to people, so problems that are detected are not necessarily resolved, and interventions are rarely tested before they touch operations. Three gaps follow. Predictions are produced for single cases, while managers decide for whole processes with shared resources, so a prediction has to reach a scheduler or an allocation algorithm to be of use. Actions interact: reprioritising orders and reallocating staff compete for the same people and the same time, and an update to the information system changes which objects and steps are affected before anyone sees the effect. And whether an action helped is only known afterwards, which makes simulation models and digital twins of organizations necessary to test interventions first. Action-oriented process mining, the loop from monitored constraints to executed actions, was the topic of my doctoral thesis at RWTH Aachen University (2019 to 2024, Best Dissertation Award Runner-Up at BPM 2024). The predictive groundwork goes back to my master's thesis at POSTECH on predicting process performance with deep neural networks. ## Research questions The line has three pillars. **Predictive process monitoring that decisions can use.** A prediction is only valuable if something can be done with it. - How can performance be predicted for a whole process rather than for one running case, and how can prediction be kept tractable on large event logs? - When a prediction is uncertain, how should the downstream allocation treat it, and how can the decision logic behind a prediction be explained to the people who act on it? **From insights to actions.** Monitoring results have to become management actions. - How can a stream of constraint violations be turned into a stream of management actions, and how are conflicting actions planned? - Which recommended actions are effective, compliant with the rules of the domain, fair towards the people in the process, and realistic given the information available at decision time? **Simulation and digital twins of organizations.** Interventions should be tested before they reach operations. - How can simulation models be discovered that stay faithful as the process evolves, remain transparent and configurable for what-if questions, and connect to the real ERP system? - How can a digital twin of an organization expose the control parameters of its information system, so that actions and system updates can be analysed for their structural, operational and performance impact before and after they happen? ## Contributions The contributions so far fall into six groups, each linked to the publication in which it appears. **Action-oriented process mining.** - A general framework with a constraint monitor that turns an event stream into a constraint instance stream and an action engine that turns it into transactions, formalized over multiple object classes with process-level valves, instance-level adaptors and alerts, and implemented as a ProM plug-in ([BPM Workshops 2020](/publication/2021-bpm/)). - A cube-based action engine with constraint cubes, OLAP operations and transaction mappings, evaluated by driving a real SAP ERP system with simulated logs through a gateway ([Progress in Artificial Intelligence, 2022](/publication/2022-prai/)). - A pattern-based action engine that mines temporal patterns of process problems with Allen's interval relations, maps them to actions through action graphs and schedules conflict-free action plans by makespan-minimizing scheduling ([Computers in Industry, 2023](/publication/2023-comind-pattern-based-action-engine/)). - Operational Problem Graphs and the ProAct monitoring engine for operational problems in object-centric processes ([Computers in Industry, 2025](/publication/2024-comind/)). **Digital twins of organizations.** - The Digital Twin Interface Model: an object-centric Petri net extended with valves and guards, with an operational view and a control view, realized in the dtween web service ([ICPM 2021](/publication/2021-icpm/)). - Analysis of process-aware information system updates as configuration changes of the digital twin, with structural, operational and performance impact types ([RCIS 2022](/publication/2022-rcis/)), extended into a four-step impact analysis method with prior and posterior analyses and the ProAct Impact Analysis prototype ([Information Systems, 2026](/publication/2026-is-processawareimpact/)). **Predictive process monitoring.** - Process-level performance prediction with annotated transition systems, two process representation matrices and CNN, LSTM and LRCN models adapted from traffic forecasting ([Decision Support Systems, 2020](/publication/2020-dss/)). - Next best actions instead of next most likely events, a research-in-progress technique whose usefulness was evaluated with process participants ([Wirtschaftsinformatik 2020](/publication/2020-wi/)). - Explainable predictive decision mining: decision models at the decision points of a Petri net with local and global SHAP explanations, in the eXdpn web application ([ICSOC Workshops 2022](/publication/2022-icsoc-workshop/)). - Variant-preserving event log sampling with unique, logarithmic and division strategies that cuts feature extraction and training time while preserving accuracy ([ICPM Workshops 2021](/publication/2021-ml4pm/), [Journal of Intelligent Information Systems, 2023](/publication/2023-sampling-jiis/)). - Object-centric process mining to improve predictive process monitoring: process executions kept as event graphs, as graph neural network input or as graph embeddings, for next activity, next timestamp and remaining time prediction ([Engineering Applications of Artificial Intelligence, 2023](/publication/2023-eaai/), [ECIS 2024](/publication/2024-ecis/)). - A pipeline that converts customer-support email threads into explainable ticket-escalation predictions with large language models and TreeSHAP, and withholds a prediction when a missing feature carries high impact ([BPM 2026](/publication/2026-bpm-ticketescalation/)). **Prediction-based resource allocation.** - Resource allocation as non-clairvoyant online job shop scheduling: a multi-task LSTM predicts processing time and next activity, and a minimum cost maximum flow algorithm assigns resources, including pseudo-assignments for predicted work items ([ICPM 2019](/publication/2019-icpm/)). - An uncertainty-aware version with Bayesian neural networks and Monte Carlo dropout that discards unreliable predictions, evaluated on a real tertiary-hospital emergency department log ([IEEE Access, 2023](/publication/2023-access/)). **Prescriptive recommendations.** - NeSyPA, which combines a Transformer prescriptor with Declare constraints compiled into automata and distils feasibility-masked, outcome-weighted targets into the model, with the Effective Compliance Rate metric ([BPM 2026, Best Paper Award](/publication/2026-bpm-nesyprescriptive/)). - Fair-LTN: counterfactual event-log generation by causal discovery and a Logic Tensor Network framework that optimizes KPI accuracy and counterfactual fairness jointly, with the CF-MSE and CF-RECC metrics ([BPM 2026](/publication/2026-bpm-fairness/)). - Behavioural recommendations mined from event logs as soft constraints for AI planning next to hard rules, solved by integer linear programming ([CAiSE Forum 2024](/publication/2024-caise-f/)). - Educational recommendations: behavioural features from campus management data and RuleFit rules translated into guidance on course load and sequencing, on 3,190 bachelor students ([Journal of Intelligent Information Systems, 2024](/publication/2024-jiis-learningrecommendation/)), and rules extracted from event data for study planning ([ICPM Workshops 2023, Best Paper Award](/publication/2023-icpm-edupm/)). **Business process simulation.** - A framework that integrates a simulation engine with a live ERP system, with CPN Tools driving SAP ERP through Remote Function Calls ([BPMDS 2021](/publication/2021-bpmds/)). - Online discovery of simulation models over a sliding window of an event stream, with incremental control-flow discovery and Hoeffding Adaptive Trees for predictive parameters ([BPM 2025](/publication/2025-bpm-sim/)). - Decision-aware simulation models that combine a Petri net with probabilistic decision trees, discovered by alignment and per-leaf distribution fitting, with three what-if operations on the trees ([ICSOC 2025](/publication/2025-icsoc-whiteboxsim/)), and the ProSiT tool for interactive, transparent simulation ([ICSOC Demo 2025](/publication/2025-icsoc-prosit/)). --- # Process discovery and diagnostics Turning event data into process models and diagnoses that people can read and trust, from sound yet expressive discovery to performance patterns, deviations and resource effects. - Type: Research line - URL: https://www.gyunam.com/research-line/process-discovery-and-diagnostics/ ## Practical problem Before anyone can predict, recommend or simulate, they need an accurate picture of how a process actually runs. Discovered models are either sound but too permissive, because block-structured miners cannot express real branching, or faithful to branching but unsound. Variant lists and performance views drown analysts in detail: thousands of variants, dense timelines in which batching and reordering hide among noise, and deviation scores that ignore the context, such as a pandemic wave or a staff shortage, that explains them. In manufacturing, which machines a product passes through matters for yield and quality, yet single-resource statistics are confounded by the many paths through hundreds of interchangeable machines. This line develops the descriptive and diagnostic side of process mining: discovery, model decomposition, performance and deviation diagnostics, and resource analysis. It started in industry projects with Samsung Electronics (2017 to 2020) on resource mining in semiconductor manufacturing and now centres on the Partially Ordered Workflow Language. It is the counterpart of [prescriptive process analytics](/research-line/prescriptive-process-analytics/), which acts on the insight. ## Research questions The line has three pillars. **Sound yet expressive discovery.** Soundness guarantees and expressiveness have been traded against each other. - How can process models be discovered that are guaranteed sound and still express non-block-structured decisions and loops, and how can existing workflow nets be reused in such a hierarchical language? - How can large models be decomposed into coherent fragments so that conformance checking scales and diagnoses point to a recognisable subprocess? **Diagnostics.** Patterns, deviations and resource effects should be detected rather than read off by an expert. - How can temporal performance patterns such as batching and reordering, and deviations that depend on context, be detected automatically? - Which resources and resource paths explain yield and quality differences when products pass through hundreds of interchangeable machines? **Non-classical traces.** Process mining applies beyond business process logs. - Which behaviours become visible when process mining is applied to learners' click streams or to the reasoning steps of large language models? ## Contributions The contributions so far fall into five groups, each linked to the publication in which it appears. **Expressive yet sound discovery with the Partially Ordered Workflow Language.** - Choice graphs, a generalization of the exclusive-choice operator to non-block-structured decisions, the POWL 2.0 language and an Inductive Miner extension with a choice graph cut that provably preserves the fitness guarantee, in the powl Python library ([BPM 2025](/publication/2025-bpm-powl/)). - A recursive algorithm that translates safe and sound workflow nets into POWL by identifying exclusive-choice, loop and partial-order patterns, with proofs that each step preserves soundness and language and completeness on the class of semi-block-structured nets ([Petri Nets 2025](/publication/2025-petri/)). - A robust inductive discovery algorithm for POWL 2.0 with several mining strategies for choice graphs that range in strictness to resolve the ambiguity between genuine loops and hidden concurrency ([Process Science, 2026](/publication/2026-procsci-sounddiscovery/)). **Decomposition and conformance.** - Isolated subnets of a workflow net, an integer linear programming formulation that extracts them around a seed transition into a hierarchical decomposition, proofs that the sub-workflow nets inherit soundness and free-choiceness, and a construction of valid decompositions for divide-and-conquer conformance checking ([CAiSE 2026](/publication/2026-caise-decomposition/)). - Conformance checking and performance analysis on object-centric directly-follows graphs, aligning per object type and mapping moves back onto the graph, in the Explori tool ([BPM Forum 2024](/publication/2024-bpm-forum/)). **Performance and deviation diagnostics.** - PerSpect, an interactive tool that reimplements the Performance Spectrum as the basis of a pattern-oriented workflow, with automated batch detection by DBSCAN clustering, propagating filters and side-by-side comparison of spectra ([CAiSE Forum 2026](/publication/2026-caise-forum-perspect/)). - Context-aware deviation detection as a four-class labeling problem, a context ontology with intrinsic, relational and external contexts, and a post-processing framework that plugs in any score-based detector, in the contect web service ([BPM Forum 2022](/publication/2022-bpm-forum/)). **Resource analysis in manufacturing.** - Resource-oriented transition systems for yield enhancement: a finite state machine whose states are the machines of each process step, overlaid with best- and worst-case wafer paths and simplified by statistical tests, evaluated on data from one of the largest Korean semiconductor manufacturers ([IEEE Transactions on Semiconductor Manufacturing, 2021](/publication/2020-ieee-tsm/)). - Quality-aware resource model discovery, a five-step method that ranks resources as superior, inferior or moderate from case-annotated transition systems, with a Python tool for interactive analysis ([Applied Sciences, 2021](/publication/2021-applied-sciences/)). **Process mining on non-classical traces.** - A taxonomy of nine learning cohorts derived from engagement, direction, repetitiveness and orderliness, with learners mapped to cohorts by dynamic time warping clustering and trend tests on 18,405 learners of an online education platform ([Education and Information Technologies, 2024](/publication/2024-kim-unlocking/)). - PMLRM-Bench and a reasoning analysis pipeline that classifies the chain-of-thought steps of 17 large reasoning models by type and effect, with a case study that reconfigures a model through system prompts ([BPM Workshops 2025](/publication/2025-bpm-workshop-lrm/)). --- ## Funded research/industry projects (11) --- # euroFMX: Foundational Models for Trustworthy Industrial Generative AI in Europe EU project building trustworthy foundation models for industrial generative AI; the TU/e contribution covers graph-based data modelling, neuro-symbolic AI and active inference for AI agents in manufacturing. - Type: Funded project - URL: https://www.gyunam.com/funding/eurofmx/ - Period: 2026 – 2029 - Kind: research - Funder: European Commission - Grant number: 101299128 - Role: Co-PI; co-author of the proposal and co-lead of the TU/e contribution - Total budget: EUR 45 M - Partners: 72 partners from 20 countries, 10 industrial pilots - Project website: https://eurofmx.eu - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md), [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract euroFMX develops foundation models for trustworthy industrial generative AI in Europe, with 72 partners from 20 countries and 10 industrial pilots. The Eindhoven University of Technology contribution, which I co-lead, covers graph-based data modelling, neuro-symbolic AI and active inference for AI agents in manufacturing. ## Objectives - Graph-based data modelling of manufacturing processes, building on object-centric event data. - Neuro-symbolic AI that combines foundation models with symbolic process models. - Active inference for AI agents that act in manufacturing processes. --- # Auditing Process Improvement Industry collaboration with EY on a tool for auditing processes. - Type: Industry project - URL: https://www.gyunam.com/funding/ey-auditing/ - Period: 2026 – present - Kind: industry - Funder: EY (Ernst & Young) ## Abstract Industry collaboration with EY at Eindhoven University of Technology: development of a tool for EY's auditing processes. ## Outputs - One conference paper submitted. --- # PMTk: Process Mining Toolkit Development of a process-mining software toolkit for Process Intelligence Solutions. - Type: Industry project - URL: https://www.gyunam.com/funding/pmtk/ - Period: 2025 – 2026 - Kind: industry - Funder: Process Intelligence Solutions GmbH - Role: PI; negotiated and lead the contract ## Abstract Industry contract with Process Intelligence Solutions GmbH for the development of a process-mining software toolkit, carried out at Fraunhofer FIT. --- # JAIF: JUPITER AI Factory EuroHPC AI Factory around JUPITER, Europe's first exascale supercomputer; Fraunhofer FIT task supporting start-ups, SMEs, industry and academia in using JUPITER for AI. - Type: Funded project - URL: https://www.gyunam.com/funding/jaif/ - Period: 2025 – 2028 - Kind: research - Funder: EuroHPC Joint Undertaking, German Federal Ministry of Education and Research (BMBF), and the science ministries of North Rhine-Westphalia and Hesse - Role: Subtask lead; responsible for a Fraunhofer FIT task - Total budget: EUR 55 M - Project website: https://www.jaif.ai ## Abstract The JUPITER AI Factory makes JUPITER, Europe's first exascale supercomputer, usable for AI. I am responsible for a Fraunhofer FIT task that supports start-ups, SMEs, industry and academia in using JUPITER for AI. --- # AI-BioDynamics: AI for Scalable Biodiversity Monitoring Bilateral German-Korean project building a biodiversity monitoring platform that combines satellite, radar and citizen-science data with knowledge-guided deep learning, transfer learning and process mining. - Type: Funded project - URL: https://www.gyunam.com/funding/ai-biodynamics/ - Period: 2025 – 2029 - Kind: research - Funder: German Federal Ministry of Education and Research (BMBF) and Korean Ministry of Science and ICT (MSIT), bilateral DE-KR programme - Role: PI (German side); author of the German proposal and lead of the German consortium - Total budget: EUR 1.3 M - Partners: Fraunhofer FIT, Geo Engine GmbH, Seoul National University of Science and Technology, Kwangwoon University, Moonsoft Co. Ltd. - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract AI-BioDynamics builds a monitoring platform for biodiversity at scale. It combines satellite, radar and citizen-science data with knowledge-guided deep learning, transfer learning and process mining. I lead the German consortium at Fraunhofer FIT; the Korean partners are Seoul National University of Science and Technology, Kwangwoon University and Moonsoft, with Geo Engine GmbH on the German side. --- # Process Modeling Using LLMs Fraunhofer-funded project on generating process models with large language models. - Type: Funded project - URL: https://www.gyunam.com/funding/process-modeling-llms/ - Period: 2024 – 2025 - Kind: research - Funder: Fraunhofer ZV - Total budget: EUR 30 k - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract Internal Fraunhofer project on process modeling with large language models, carried out at Fraunhofer FIT. --- # PM4KNIME: Process Mining Extension for KNIME Design and development of an open process-mining extension for the KNIME analytics platform. - Type: Industry project - URL: https://www.gyunam.com/funding/pm4knime/ - Period: 2024 – present - Kind: industry - Funder: KNIME AG - Role: PI; negotiated and lead the contract ## Abstract Industry contract with KNIME AG for the design and development of an open process-mining extension for the KNIME analytics platform, carried out at Fraunhofer FIT. --- # AIStudyBuddy This project aims to introduce AI-driven resources and assessments to enhance the teaching and learning experiences for students and curriculum planners. The AI-enabled 'study companion' tool offers targeted assistance to students, like setting objectives to secure top grades in a course and generating suggestions to meet those objectives. The AI-integrated 'companion analytics' tool equips curriculum planners with dashboards and assists them in the process of (re)structuring study programs. - Type: Funded project - URL: https://www.gyunam.com/funding/aistudybuddy/ - Period: 2023 – 2024 - Kind: research - Funder: German Federal Ministry of Education and Research (BMBF), project agency VDI/VDE-IT - Role: Led the RWTH Aachen University team in two work packages on AI-based study support for students - Total budget: EUR 3.9 M - Project website: https://learntech.rwth-aachen.de/cms/LearnTech/Forschung/Projekte/~siepo/AIStudyBuddy/?lidx=1 - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract AIStudyBuddy introduces AI-driven resources and assessments to improve teaching and learning for students and curriculum planners. The AI-enabled study companion gives students targeted assistance, such as setting objectives for a course and generating suggestions to meet them. The companion analytics tool gives curriculum planners dashboards and support for (re)structuring study programmes. ## Outputs - [Learning recommendations from educational event data in higher education](/publication/2024-jiis-learningrecommendation/) - [Extracting rules from event data for study planning](/publication/2023-icpm-edupm/) - [Unlocking learner engagement and performance](/publication/2024-kim-unlocking/) --- # Root cause analysis of bottlenecks in FAB - Type: Industry project - URL: https://www.gyunam.com/funding/samsung-3rd/ - Period: 2019 – 2020 - Kind: industry - Funder: Samsung Electronics - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) --- # Development of algorithm for recommending best resource path using AI - Type: Industry project - URL: https://www.gyunam.com/funding/samsung-2nd/ - Period: 2018 - Kind: industry - Funder: Samsung Electronics - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) - Tags: manufacturing, samsung, aim --- # Development of best reference resource mining algorithm The importance of efficiently operating and managing manufacturing equipment in a process is highlighted in this research. By utilizing process mining techniques, analysis of equipment status and operation status becomes possible. The research aims to develop an equipment mining algorithm for deriving the Best Reference equipment. The study involves analyzing core semiconductor manufacturing processes, identifying problem equipment, engaging in process mining technology sensing activities that include technology and case introductions, and developing a methodology for identifying the Best Reference equipment using process mining techniques. - Type: Industry project - URL: https://www.gyunam.com/funding/samsung-1st/ - Period: 2017 - Kind: industry - Funder: Samsung Electronics - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) - Tags: manufacturing, samsung, aim --- ## Software (2) --- # ProAct: Action-Oriented Process Mining *A Tool for Action-Oriented Process Mining: Turning Events to Actions.* Action-oriented process mining focuses on generating necessary management actions that enhance business processes' performance. These actions are derived by analyzing event data through process mining techniques. ProAct is a tool that supports action-oriented process mining. ProAct is comprised of three main components: *ProAct: Monitoring Engine*, *ProAct: Action Engine*, *ProAct: Impact Analysis*. - Type: Software - URL: https://www.gyunam.com/software_project/proact/ - Link (Document): https://proact.readthedocs.io/ - Link (Source Code): https://github.com/gyunamister/proact --- # OCPA: Object-Centric Process Analysis A Python library dedicated to the analysis of object-centric event data. The library includes capabilities for handling and processing object-centric event logs, carrying out process discovery, assessing process models, analyzing process performance, and supporting predictive process monitoring. This project is a collaboration with [Niklas Adams](https://niklasadams.com). - Type: Software - URL: https://www.gyunam.com/software_project/ocpa/ - Link (Document): https://ocpa.readthedocs.io/ - Link (Source Code): https://github.com/ocpm/ocpa --- ## Publications (60) --- # Toward Automated Ticket Escalation in Customer Support Processes: From Emails to Explainable Predictions - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-bpm-ticketescalation/ - Date: 2026-09-27 - Authors: Furkan Kocak, Gyunam Park - Venue: Business Process Management - DOI: https://doi.org/10.1007/978-3-032-37777-7_27 - PDF: https://www.gyunam.com/publication/2026-bpm-ticketescalation/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Business processes in customer support often remain undocumented, with critical decisions such as ticket escalation relying on tacit knowledge rather than explicit process models. Automating such process decisions is challenging because the underlying data is unstructured email text, not structured event logs assumed by existing process mining and predictive monitoring approaches. We present a novel pipeline artifact that addresses this gap: a semi-automated four-stage pipeline that extracts structured features from customer support emails using large language models (LLMs) together with lightweight domain-expert validation, and employs explainable machine learning to predict the optimal timing of ticket creation, a key decision point in the support process. Our pipeline generates an initial feature schema from sample threads, consolidates semantically similar features through embedding-based clustering under domain-expert review, extracts cumulative feature values at each email point, and constructs a structured dataset ready for machine learning. We introduce a SHAP-based active feature learning strategy that identifies which missing features, when acquired, would most improve prediction confidence, enabling principled decisions about when to request additional information versus making immediate predictions. We evaluate our approach in two stages: first, on a reproducible synthetic dataset where comparison against six baselines demonstrates that LLM-extracted features provide complementary semantic information, with combined features achieving the best overall performance (ROC-AUC 0.90) while end-to-end LLM classification performs worst (ROC-AUC 0.46), validating our design of using LLMs as feature extractors rather than decision makers; second, on 1,500 email instances from 800 real-world email threads from Nosco, a commercial CRM platform, where we demonstrate effective feature extraction, predictive performance, and active feature learning for practical deployment. ## Practical problem Customer support teams, especially in small and medium-sized businesses, decide when an email conversation should become a formal support ticket, and that judgment lives in the heads of experienced agents rather than in documented rules. The raw material is unstructured email text, so predictive and prescriptive monitoring techniques for business processes, which expect event logs with case identifiers, activities and timestamps, cannot be applied directly. Earlier ticket-escalation predictors needed analysts to hand-engineer features from interviews, an effort such businesses cannot afford, while letting a large language model make the decision itself raises trust, auditability and privacy concerns. Support organizations need a way to turn email threads into structured, explainable predictions and to know when more information should be requested before deciding. ## Research questions 1. How can unstructured customer support emails be transformed, with bounded domain-expert effort, into a structured dataset for predicting when a ticket should be created? 2. Do features extracted by large language models improve escalation-timing prediction compared with lexical, structural, rule-based and end-to-end language-model baselines? 3. Can SHAP-based reasoning about missing features tell the system when to withhold a prediction and request additional information instead? ## Contributions - A semi-automated four-stage pipeline (feature identification, semantic clustering and consolidation, feature extraction, dataset construction) that converts email threads into cumulative per-email feature vectors using large language models (LLMs). - An embedding-based consolidation step (mxbai-embed-large-v1 embeddings, HDBSCAN density-based clustering) in which domain experts pick representative features, merge correlated ones and bound categorical value sets. - An explainable gradient-boosted classifier with TreeSHAP (SHapley Additive exPlanations) attributions that predicts ticket-creation timing at each email point, including attributions for missing features. - A SHAP-based active feature learning algorithm that withholds a prediction when missing features carry high impact or make the prediction unstable, and names the feature to request. - An evaluation on a public synthetic dataset of 350 threads and on 1,500 email instances from 800 threads of the Nosco customer relationship management platform, with code released on Zenodo. ## Key findings - On the synthetic dataset, combining LLM-extracted, metadata and term-frequency features with CATBoost reaches an area under the receiver operating characteristic curve (ROC-AUC) of 0.90 ± 0.05 and F1 0.85 ± 0.02, adding 0.13 AUC over metadata-only features. - End-to-end LLM classification performs worst (ROC-AUC 0.46 zero-shot, 0.48 few-shot), supporting the design of using LLMs as feature extractors rather than decision makers. - On Nosco data, consolidation reduces 785 LLM-generated features to 35 (95.5% reduction) in about four hours of expert effort; CATBoost achieves ROC-AUC 0.73 ± 0.04 and F1 0.67 ± 0.04, the best of six classifiers. - Active feature learning on 150 test instances raises precision from 0.67 to 0.74 and ROC-AUC from 0.73 to 0.75 while predicting on 50.7% of instances; recall drops from 0.67 to 0.61. - Limitations: the real-world set contains only threads that ended in a ticket, extraction varies across models and prompts, threads are treated independently, and the Nosco data cannot be released. --- # Neuro-Symbolic Prescriptive Process Analytics: Learning Compliant and Effective Process Recommendations Through Constraint Distillation - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-bpm-nesyprescriptive/ - Date: 2026-09-27 - Authors: Gyunam Park - Venue: Business Process Management - DOI: https://doi.org/10.1007/978-3-032-37777-7_30 - PDF: https://www.gyunam.com/publication/2026-bpm-nesyprescriptive/paper.pdf - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md), [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Given a running process case, prescriptive process analytics (PrPA) recommends the next activity that best improves a desired outcome, such as reducing cycle time. Recommendations must also be operationally feasible: an activity that improves the predicted outcome but violates process constraints cannot be enacted. The core challenge is therefore not compliance alone, which inference-time filtering can enforce, but achieving high outcome quality within the compliant space. Existing approaches fall short: post-hoc filtering never learns which compliant activities are outcome-improving, while logic-as-loss penalties balance unstable competing gradients from data fit and compliance. Both also treat all constraints equally, so combining mandatory hard constraints with relaxable soft constraints can leave many cases with no feasible recommendation. We propose NeSyPA, a neurosymbolic PrPA method with two contributions: Hierarchical Constraint Relaxation, which drops soft constraints only when the full constraint set is unsatisfiable, and Constraint-Compiled Outcome Distillation, which compiles per-prefix feasible sets into feasibility-masked, outcome-weighted soft targets that teach the neural prescriptor to rank feasible activities by outcome merit. On four real-world event logs, NeSyPA achieves the highest Effective Compliance Rate on all test sets, with gains of +6.7% to +9.9%, driven by higher outcome quality within the compliant space. ## Practical problem Case workers handling a running loan application or a traffic fine want a system that tells them which step to take next so the case ends well, for instance with an accepted offer or a fine paid on time. Such a recommendation is useless if it breaks the rules of the process: skipping a mandatory review may look good for cycle time but can never be enacted. Today's recommenders either train without rules and filter out forbidden steps afterwards, so they never learn which allowed steps actually help, or they add a rule-violation penalty to training, which forces an unstable balancing act between fitting data and obeying rules. On top of that, real rule books mix regulations that must never be broken with habitual patterns mined from history, and when both are enforced equally, a large share of cases ends up with no permitted next step at all. ## Research questions 1. How can a neural recommender be trained to rank only constraint-compliant next activities by their expected outcome improvement, without a tunable trade-off between compliance and outcome quality? 2. How should mandatory (hard) and relaxable (soft) process constraints be prioritized so that cases with over-constrained knowledge bases still receive a feasible recommendation? ## Contributions - NeSyPA, a neurosymbolic prescriptive process analytics method combining a Transformer prescriptor with Declare constraints compiled into deterministic finite automata. - Hierarchical Constraint Relaxation (HCR), which partitions the knowledge base into hard and soft constraints and drops the soft tier only when the full set admits no activity. - Constraint-Compiled Outcome Distillation (CCOD), which turns per-prefix feasible sets into feasibility-masked, outcome-weighted soft targets and fine-tunes the activity head in one pass. - Effective Compliance Rate (ECR), a metric counting recommendations that are both outcome-improving and constraint-compliant, alongside outcome-improvement and constraint-satisfaction rates. - An evaluation protocol with four baselines (unconstrained, post-hoc filtering, reinforcement learning, logic-as-loss) on four real-life test sets, with code and data on GitHub (gyunamister/nesy_prpa). ## Key findings - NeSyPA reaches the highest Effective Compliance Rate on all four test sets: 0.569 on BPIC 2012, 0.437 and 0.456 on the two BPIC 2017 splits, 0.399 on Traffic Fines, i.e. +6.7, +9.7, +9.9 and +0.03 percentage points over the best baseline. - The gains come from outcome quality within the compliant space: Outcome Improvement Rate rises to 0.849, 0.718 and 0.735 on the BPIC logs versus 0.505 to 0.711 for baselines, while constraint satisfaction stays comparable (about 0.60 to 0.68). - In the ablation, CCOD is the dominant component (+7.1, +11.5, +10.0 and +1.5 points over the post-hoc variant with hierarchy); hierarchy adds at most +1.1 points but cuts irreparable prefixes from 31.8% to 2.3% on BPIC 2012. - The logic-as-loss baseline degrades constraint satisfaction to 0.753 on Traffic Fines against 0.940 for filtering baselines, and the reinforcement-learning baseline shows high variance (coefficient of variation 0.33 on BPIC 2017 before). - Injecting 10% spurious soft constraints on BPIC 2012 drops ECR from 0.574 to about 0.345 and then plateaus; relaxation rate and ECR degradation correlate at r = -0.91, and the all-or-nothing soft relaxation remains a limitation. --- # Neuro-symbolic AI for BPM - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-bpm-nesybpm/ - Date: 2026-09-27 - Authors: Gyunam Park, Fabrizio De Santis - Venue: Business Process Management - DOI: https://doi.org/10.1007/978-3-032-37777-7_4 - PDF: https://www.gyunam.com/publication/2026-bpm-nesybpm/paper.pdf - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract Process analytics has become predominantly data-driven: deep sequence models such as LSTMs and Transformers now dominate predictive and prescriptive process monitoring and deliver strong average-case accuracy. These models are, however, opaque, require large amounts of data, and do not incorporate the process knowledge that organizations already hold in the form of regulations, declarative constraints, and reference process models. The symbolic side of Business Process Management (BPM) carries exactly this knowledge but is brittle on noisy, real-world event data. Neuro-symbolic AI (NeSy) aims to bridge these two traditions, and a growing body of BPM work has begun to use it. This tutorial gives the BPM community an accessible overview of the relationship between BPM and NeSy. It introduces Kautz's architectural taxonomy of NeSy and the Learning from Logical Constraints paradigm together with its main implementation, Logic Tensor Networks. It then traces how NeSy attaches to concrete BPM analytics tasks through a sequence of case studies spanning compliance-aware prediction, robust rule injection, healthcare outcome prediction, and prescriptive analytics. The tutorial highlights the failure modes that practitioners must avoid when injecting domain knowledge into deep process models, examines how to evaluate whether injected knowledge actually helps, and closes with open problems and a practical decision aid for when NeSy is worth adopting. ## Practical problem Organizations that run predictive or prescriptive process analytics rely on deep sequence models that forecast next activities, outcomes or remaining times from event logs. These models are opaque, need large amounts of data, and ignore the regulations, declarative constraints and reference process models the organization already holds, so they can predict behavior that violates known rules, most sharply where compliant behavior is rare and the stakes are high. The symbolic side of business process management carries exactly that knowledge but breaks down on the noisy, incomplete logs that real systems produce. Practitioners need to understand when and how to combine the two traditions, and how to avoid injecting knowledge in ways that make a model worse than having no knowledge at all. ## Research questions 1. What is neuro-symbolic artificial intelligence and where does it attach to the predictive and prescriptive process analytics pipeline? 2. How does injecting domain knowledge into deep process models help, and how and when does it fail? 3. How can practitioners evaluate whether injected knowledge actually helps and decide whether a neuro-symbolic approach is worth adopting? ## Contributions - A tutorial that organizes neuro-symbolic work in business process management (BPM) using Kautz's six-type architectural taxonomy and the Learning from Logical Constraints paradigm. - An introduction to Logic Tensor Networks and their Real Logic formalism sufficient to read and write first-order axioms that compile into training signals for deep models. - A sequence of case studies spanning compliance-aware prediction, two-stage training with rule pruning, intensive-care mortality prediction, prescriptive analytics (NeSyPA) and neuro-symbolic anomaly detection. - An analysis of failure modes and open problems: vacuous satisfaction, the soft-compliance gap, sequential constraints under fuzzy logic, and the conditions under which knowledge utility is observable. - A five-question decision aid (data scarcity, rule availability, regulatory exposure, tolerable opacity, deployment horizon) for judging when neuro-symbolic methods pay off. ## Key findings - The tutorial reports no new experiments; it synthesizes published results in which output refinement improves F1 by up to 18.4% on Sepsis and 10.1% on the Business Process Intelligence Challenge 2012 log with only 4.55% compliant training traces. - It shows that naive logic losses can backfire: an unguarded Logic Tensor Network drops F1 to 18.79% on Sepsis through vacuous satisfaction, and two-stage training with analytical rule pruning recovers 75.82%. - It argues that Type-4 integration yields training-time soft compliance rather than runtime guarantees, that fuzzy aggregation handles atemporal but not sequential constraints, and that knowledge pays off mainly when data is scarce and reliable rules exist. --- # Fairness in Prescriptive Process Analytics: Integrating Fairness Constraints into the Learning Phase - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-bpm-fairness/ - Date: 2026-09-27 - Authors: Fabrizio De Santis, Alessandro Padella, Gyunam Park, Francesco Zanichelli, Massimiliano de Leoni - Venue: Business Process Management - DOI: https://doi.org/10.1007/978-3-032-37777-7_23 - PDF: https://www.gyunam.com/publication/2026-bpm-fairness/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Prescriptive Process Analytics is a core field in business process management that provides automated activity recommendations to improve Key Performance Indicators in business processes. Because such recommendations can directly affect individuals involved in the process, ensuring fairness is essential. Recent work has begun to address fairness in process monitoring by enforcing group-level parity across demographic groups at the predictive stage. However, these approaches leave the prescriptive phase entirely unexamined and rely solely on group fairness notions, failing to account for individual-level counterfactual fairness (i.e., the requirement that a recommendation should not change when sensitive attributes are altered). To address this gap, we propose to enforce counterfactual fairness directly in prescriptive process monitoring. We develop a novel neuro-symbolic framework based on Logic Tensor Networks that encodes counterfactual fairness as a logical axiom over predictions, seamlessly integrated into the learning process. We evaluate our approach on real-world event logs and find that it enforces counterfactual fairness at the prediction level and improves recommendation-level counterfactual consistency under the considered policy, while maintaining competitive predictive and prescriptive performance, establishing the first fairness-aware framework for Prescriptive Process Analytics. ## Practical problem Companies increasingly let software recommend the next step in a running case, such as whether a loan application goes to manual review or a fast track, or which department a patient is sent to. Those recommendations are learned from historical data, so they can quietly depend on gender, age, or nationality, and the person affected has no way of knowing. The fairness tools that exist for process analytics work only on the prediction side and only check that groups look equal on average, which still allows an individual to receive a different recommendation purely because of a protected attribute. Process owners and compliance teams therefore lack a way to guarantee that a given case would be treated the same if only its sensitive attribute were different. ## Research questions 1. How can counterfactual fairness, invariance of predictions and recommendations under a change of the sensitive attribute, be enforced in prescriptive process analytics rather than only in predictive monitoring? 2. Does enforcing counterfactual fairness at the prediction level carry over to the recommendation level, and at what cost in predictive accuracy and recommendation effectiveness? ## Contributions - A counterfactual event-log generation procedure that uses Fast Causal Inference (FCI) to discover causal dependencies among case-level attributes and fits structural equations to intervene on a sensitive attribute consistently. - A neuro-symbolic prescriptive framework (Fair-LTN) built on Logic Tensor Networks that jointly optimizes a regression axiom for key performance indicator (KPI) accuracy and a counterfactual fairness axiom. - Two evaluation metrics for fairness in process analytics: Counterfactual Mean Squared Error (CF-MSE) at prediction level and Counterfactual Recommendation Score (CF-RECC) at recommendation level. - A simulation-based evaluation protocol using the `ProSiT` simulator on original and counterfactual test logs, with code released on GitHub (`FabrizioDeSantis/Fair-PrescriptiveProcessMonitoring`). - An empirical study on BPI Challenge 2012, three Hospital Billing scenarios, and BPI Challenge 2013 incidents with gender, age, and organization country as sensitive attributes. ## Key findings - Fair-LTN keeps or improves predictive accuracy: on BPI Challenge 2012 its mean absolute error for remaining time is 12.46 days versus 13.08 for a plain LSTM and 12.64 for an LTN without the fairness axiom. - Prediction-level fairness improves sharply: CF-MSE on Hospital Billing with gender bias falls from 41.53 (LSTM) and 42.20 (LTN-Data) to 2.85, and on BPI Challenge 2013 from 265.76 and 199.71 to 11.83. - Where the sensitive attribute changes the ranking of candidate activities, recommendation consistency recovers: CF-RECC rises from 0.77 to 0.94 on Hospital Billing with gender bias and from 0.68 to 0.85 on BPI Challenge 2013. - Fairness costs little effectiveness: simulated trace durations under Fair-LTN stay within 0.3 days of the unfair baselines on the BPI logs and within 1.7 days on Hospital Billing, where all methods cut duration by 33-56% and 3-7% respectively. - The carry-over to prescriptive fairness holds only for a policy ranking activities by predicted KPI with a fair transition system; where few candidate activities exist, unfair baselines already exceed 0.95 CF-RECC, and results depend on the discovered causal graph. --- # Analyzing structural, operational, and performance impact of process-aware information system updates - Type: Journal article - URL: https://www.gyunam.com/publication/2026-is-processawareimpact/ - Date: 2026-08-20 - Authors: Gyunam Park, Marco Comuzzi, Wil M. P. van der Aalst - Venue: Information Systems - DOI: https://doi.org/10.1016/j.is.2026.102794 - PDF: https://www.gyunam.com/publication/2026-is-processawareimpact/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Digital transformation often entails small-scale updates to the business functions exposed by enterprise systems, which support the execution of organizational business processes. These changes may increase operational friction in process execution, decreasing the process performance, increasing the complexity of maintenance activities, and making system documentation and training obsolete. Detecting and understanding their impact on process execution is therefore crucial to maintain smooth operations. Previous research contributions concerning the tracking and impact analysis of information systems updates are limited in scope, functionality, and tool support. In this paper, we propose a data-driven approach that uses the recently developed Digital Twins of Organizations (DTOs) to assess the impact of (process-aware) information systems updates. We model the updates using the configuration of DTOs and the newly introduced concept of actions in DTOs. The impact of the application of actions in a given configuration is quantitatively assessed by different types of impacts: structural, operational, and performance-related. We show the feasibility of the proposed approach by discussing a software prototype implementing it, a case study involving artificially-generated data from a standard ERP procure-to-pay process, and a case study using real-world data from a loan application process. ## Practical problem Companies continuously tweak the enterprise systems (ERP, CRM) that run their business processes, for example by raising the minimum order quantity in a purchasing function or changing when a loan application is cancelled. Each small rule change ripples through the process: upstream and downstream functions change behavior, orders already in flight are caught by the new rule, and waiting and throughput times shift. IT and process owners today have little support for seeing these effects, because existing change-management approaches reason only about the model (which functions are connected) and ignore what the recorded execution data says about running instances and performance. Without a data-driven view, updates are rolled out blind and their side effects are discovered only when operations slow down or staff and documentation fall out of step. ## Research questions 1. Given a process-aware information system and an update to the business rules of one of its functions, how can the structural, operational, and performance impacts of that update be systematically identified and quantified from recorded execution data? 2. How can a digital twin of an organization, built on object-centric Petri nets, be used to model both the system and its updates so that such impacts become computable? 3. Is the resulting method feasible to implement and useful on realistic enterprise data, both simulated and real? ## Contributions - A meta-model of process-aware information systems that relates business processes, functions, objects, constraint and operation rules, updates, and three impact types (structural, operational, performance). - Formalizes the Digital Twin Interface Model (DT-IM): an object-centric Petri net with linear rule expressions, configurations, operational states, interface bindings, and actions that model system updates as configuration changes. - A four-step impact analysis method: change detection, prior and posterior structural impact analysis, replay-based prior and posterior operational impact analysis, and function-wise and object-wise performance impact analysis. - A web-based prototype, `ProAct: Impact Analysis`, with open source code and a user manual, implementing the four analyses with interactive visualizations. - Two scenario-based case studies: a simulated procure-to-pay system mirroring SAP ERP rules, and a real-life loan application process (BPI Challenge 2017). ## Key findings - In the simulated procure-to-pay case, raising the minimum material quantity for create purchase order from 5 to 10 for five days affects 37 in-flight purchase requisitions and 95 materials upstream, and up to 32 purchase orders, 24 goods receipts, and 10 invoices before clear invoice. - The same action reduces the average sojourn time of create purchase order by 9 minutes, cuts the number of purchase orders by 13 during the change window, and shortens average purchase order throughput time by 4 minutes. - On the real loan application log, five consecutive two-week rule changes to cancel application affect 547, 534, 529, 489, and 410 open applications respectively; 517 of the 547 in the first window sit after create offer, pinpointing where to allocate resources. - Operational impact counts expose process structure: applications affected after create versus submit differ by only 2 (547 vs 545), and for the call after offers action the offers waiting for return offer equal those waiting for cancel application (321). - Limitations: the structural analysis treats all connected transitions as equally impacted, operational counts are only as reliable as the alignment between model and log, and performance impacts can be measured only after the change is applied, so the method offers limited foresight. --- # A discovery technique for expressive yet sound process models - Type: Journal article - URL: https://www.gyunam.com/publication/2026-procsci-sounddiscovery/ - Date: 2026-06-10 - Authors: Humam Kourani, Gyunam Park, Wil M. P. van der Aalst - Venue: Process Science - DOI: https://doi.org/10.1007/s44311-026-00046-8 - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) ## Abstract Process discovery enables organizations to analyze and improve their operations by automatically deriving process models from event logs. While the Inductive Mining framework is widely adopted for ensuring model soundness, its strict block-structured nature often fails to capture the true complexity of real-world control flows. Recent advancements, such as the Partially Ordered Workflow Language (POWL), have relaxed these constraints for concurrency, yet a significant gap remains in effectively modeling complex decision logic and unstructured loops. We bridge this gap by introducing POWL 2.0 , an extended modeling language that integrates choice graphs to represent non-block-structured decisions and cyclic flows within a hierarchical framework. In this paper, we present a robust inductive discovery algorithm that leverages POWL 2.0, and we explore several mining strategies for choice graphs. These strategies range in strictness to resolve the ambiguity between genuine loops and hidden concurrency. Our experimental evaluation demonstrates that the proposed approach captures complex behaviors without compromising the scalability and quality guarantees of the Inductive Mining framework. --- # Neuro-Symbolic Process Anomaly Detection - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-caise-nesyanomaly/ - Date: 2026-06-09 - Authors: Devashish Gaikwad, Wil M. P. van der Aalst, Gyunam Park - Venue: Advanced Information Systems Engineering - 38th International Conference, CAiSE 2026, Verona, Italy, June 8-12, 2026, Proceedings, Part II - DOI: https://doi.org/10.1007/978-3-032-28117-3_14 - PDF: https://www.gyunam.com/publication/2026-caise-nesyanomaly/paper.pdf - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract Process anomaly detection is an important application of process mining for identifying deviations from the normal behavior of a process. Neural network-based methods have recently been applied to this task, learning directly from event logs without requiring a predefined process model. However, since anomaly detection is a purely statistical task, these models fail to incorporate human domain knowledge. As a result, rare but conformant traces are often misclassified as anomalies due to their low frequency, which limits the effectiveness of the detection process. Recent developments in the field of neuro-symbolic AI have introduced Logic Tensor Networks (LTN) as a means to integrate symbolic knowledge into neural networks using real-valued logic. In this work, we propose a neuro-symbolic approach that integrates domain knowledge into neural anomaly detection using LTN and Declare constraints. Using autoencoder models as a foundation, we encode Declare constraints as soft logical guiderails within the learning process to distinguish between anomalous and rare but conformant behavior. Evaluations on synthetic and real-world datasets demonstrate that our approach improves F1 scores even when as few as 10 conformant traces exist, and that the choice of Declare constraint and by extension human domain knowledge significantly influences performance gains. ## Practical problem Organizations screen their event logs for anomalous process executions to catch compliance breaches and inefficiencies. When no trustworthy process model exists, neural networks such as autoencoders learn what normal behaviour looks like from the log and flag traces they cannot reconstruct well. Because this is a purely statistical judgement, executions that are legitimate but rare, such as an uncommon but permitted approval path, get the same high reconstruction error as real deviations and are reported as anomalies, creating false alarms and raising the amount of data needed. Domain experts know which rare paths are allowed, but current neural detectors offer no way to feed that knowledge into the model. ## Research questions 1. How can human domain knowledge about rare but conformant behaviour be injected into a neural anomaly detector without abandoning learning from the event log? 2. Does combining an autoencoder with Declare constraints through Logic Tensor Networks improve anomaly detection on synthetic and real-world logs, and how many rare conformant traces are needed for the gain to appear? 3. How much does the choice of Declare constraint, and therefore the quality of the injected domain knowledge, influence the improvement? ## Contributions - A neuro-symbolic anomaly detection approach that fine-tunes a pretrained denoising autoencoder with Logic Tensor Networks (LTN), a framework that grounds first-order logic in real-valued fuzzy logic. - A knowledge base construction step that mines Declare constraints, keeps low-support high-confidence constraints, selects them with domain knowledge and translates them into first-order logic over positional activity predicates. - A training objective that combines reconstruction loss with the satisfiability of the constraint on constraint-satisfying traces and its violation on violating traces, reusing established threshold heuristics for detection. - An evaluation protocol on eight synthetic logs and BPIC12, BPIC13 and BPIC17 with 30 percent injected anomalies, varying the number of rare conformant traces from 10 to 300, plus an ablation over seven Declare templates. - An open-source implementation, `LTNcoder`, on GitHub. ## Key findings - On synthetic logs the approach beats the autoencoder baseline in 7 of 8 logs; on the Large log F1 moves from a baseline range of 0.27–0.36 to 0.57–0.66, and on Paper from 0.38–0.43 to 0.51–0.60, with absolute F1 between 0.23 and 0.66. - On real-world logs it wins in 2 of 3: BPIC12 F1 moves from a baseline range of 0.42–0.65 to 0.53–0.78 and BPIC13 from 0.26–0.30 to 0.35–0.41, while BPIC17 stays level at about 0.50–0.52. - The gain is already present with only 10 rare but conformant traces and stems from higher recall with a negligible drop in precision. - The ablation on Paper and BPIC12 shows that the F1 improvement and the post-training satisfiability vary strongly with the Declare template used; a succession constraint with low confidence reaches low satisfiability yet still yields a slight F1 gain. - The approach does not help on the Wide synthetic log and BPIC17, whose many variants exceed the autoencoder's capacity, and fine-tuning there can even lower F1 through lost generalization. --- # Compliance-Aware Predictive Process Monitoring: A Neuro-Symbolic Approach - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-caise-complianceppm/ - Date: 2026-06-09 - Authors: Fabrizio De Santis, Gyunam Park, Wil M. P. van der Aalst, Francesco Zanichelli - Venue: Advanced Information Systems Engineering - DOI: https://doi.org/10.1007/978-3-032-28117-3_13 - PDF: https://www.gyunam.com/publication/2026-caise-complianceppm/paper.pdf - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract Existing approaches for predictive process monitoring are sub-symbolic, meaning that they learn correlations between descriptive features and a target feature fully based on data, e.g., predicting the surgical needs of a patient based on historical events and biometrics. However, such approaches fail to incorporate domain-specific process constraints (knowledge), e.g., surgery can only be planned if the patient was released more than a week ago, limiting the adherence to compliance and providing less accurate predictions. In this paper, we present a neuro-symbolic approach for predictive process monitoring, leveraging Logic Tensor Networks (LTNs) to inject process knowledge into predictive models. The proposed approach follows a structured pipeline consisting of four key stages: 1) feature extraction; 2) rule extraction; 3) knowledge base creation; and 4) knowledge injection. Our evaluation shows that, in addition to learning the process constraints, the neuro-symbolic model also achieves better performance, demonstrating higher compliance and improved accuracy compared to baseline approaches across all compliance-aware experiments. ## Practical problem Organizations use predictive process monitoring to forecast how a running case will end, for example whether a loan application will be accepted or a sepsis patient will need intensive care. Today's predictors are deep neural networks that learn only from historical event data, so they ignore rules the organization already knows, such as regulations, clinical protocols and service-level agreements, and can issue predictions that contradict them, especially when compliant behavior is rare in the history or a rule is new. Earlier attempts to add knowledge either hard-code control-flow patterns as input features or correct outputs after training, and they rarely cover timing or data attributes. Process owners in compliance-sensitive settings need predictors that are accurate and consistent with the rules at the same time. ## Research questions 1. How does injecting domain knowledge improve predictive performance compared with purely data-driven approaches? 2. How do the three knowledge injection configurations (feature expansion, output refinement, parallel constraints) influence predictive performance? 3. How well do the models adhere to domain constraints, measured as the proportion of satisfied constraints over all applicable constraints? ## Contributions - A four-stage neuro-symbolic pipeline for predictive process monitoring: feature extraction, rule extraction, knowledge base creation, and knowledge injection through Logic Tensor Networks (LTNs). - A feature vocabulary of control-flow, temporal and payload features aligned with Declare constraints, service-level agreements and payload rules, so that every rule type maps onto features. - A translation of control-flow rules in Linear Temporal Logic, temporal rules and payload rules into first-order logic, with a categorization into class-dependent (outcome- or non-outcome-oriented) and class-independent knowledge. - Three injection mechanisms relative to the classification predicate: feature expansion at the input, output refinement on predicate outputs, and parallel constraints as auxiliary objectives. - A compliance-aware evaluation protocol on four public logs with eight architectural variants and a constraint-satisfaction metric; code released on GitHub (NeSyPPM-Compliance). ## Key findings - Knowledge-enhanced LTNs outperform Long Short-Term Memory (LSTM), Transformer, feature-encoding and semantic-loss baselines on all four logs; on Sepsis LTN-T reaches 93.15% accuracy and 92.06% F1 against 81.53% and 73.66% for the Transformer. - Gains are largest on small logs: F1 improves by 18.4% on Sepsis (Transformer to LTN-T) and 10.1% on the Business Process Intelligence Challenge 2012 log (LSTM to LTN-L), but only 1.96% on Traffic fines. - In the ablation, output refinement (B) delivers the strongest improvements, feature expansion (A) is limited, and parallel constraints (C) add slight gains only when combined with B. - The LTN variants keep high compliance when compliant training traces are scarce (4.55% in Sepsis, 13.09% in BPIC2012), whereas baselines are competitive only on BPIC2017, where 43.92% of training traces are compliant. - The improvement attenuates on large logs, where abundant data reduces the relative influence of logical rules, and each knowledge base is limited to six rules. --- # Flexible and Hierarchical Decomposition of Workflow Nets for Process Analysis - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-caise-decomposition/ - Date: 2026-06-08 - Authors: Tsung-Hao Huang, Lukas M. Jansen, Marco Pegoraro, Gyunam Park, Wil M. P. van der Aalst - Venue: Advanced Information Systems Engineering - DOI: https://doi.org/10.1007/978-3-032-28110-4_17 - PDF: https://www.gyunam.com/publication/2026-caise-decomposition/paper.pdf - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) ## Abstract Decomposition is widely used in process mining to enhance scalability, facilitate focused analysis, and mitigate the cognitive complexity of large-scale models. While both process discovery and conformance checking benefit from decomposition, existing approaches—especially those based on single-entry single-exit (SESE) fragments—impose structural constraints that limit diagnostic value. This paper proposes a complementary decomposition technique that partitions a WF-net into isolated subnets, i.e., internally coherent fragments with clear interfaces. The decomposition is hierarchical, independent of SESE assumptions, and compatible with valid system net decompositions, enabling divide-and-conquer conformance checking. Experiments show that our approach improves alignment runtime and yields fitness approximations that are closer to the true values than existing decomposition methods. ## Practical problem Analysts who check whether recorded process executions conform to a large process model face two problems: the computation becomes very slow for models with hundreds of activities and long traces, and a single global conformance score says little about where in the process the deviations occur. The standard remedy is to cut the model into fragments and check each one, but the dominant technique only recognizes fragments with exactly one entry and one exit. Real models with multi-entry or multi-exit routing therefore fall apart into trivial pieces, which neither speed things up much nor correspond to anything an analyst would recognize as a subprocess, and the per-fragment fitness values can be far too optimistic. ## Research questions 1. How can a workflow net be partitioned into internally coherent subnets with clear interfaces without assuming single-entry single-exit structure? 2. Can such a decomposition be made compatible with valid decompositions so that conformance checking can be split across subnets while giving faster runtimes and fitness estimates closer to the true value than existing decompositions? ## Contributions - Formalizes isolated subnets, transition-bordered fragments of a workflow net whose start transitions share one input place set and end transitions share one output place set, and a decomposition into disjoint isolated subnets covering all transitions. - An integer linear programming (ILP) formulation, solved iteratively with `PySCIPOpt`, that extracts the maximal non-trivial isolated subnet around a seed transition and yields a hierarchical, recursively refinable decomposition. - Proves that sub-workflow nets derived from isolated subnets inherit soundness and free-choiceness from the original net. - A construction (extended and induced valid decomposition) that adds place-induced subnets and merges fragments sharing silent transitions so the result is a valid decomposition for divide-and-conquer conformance checking. - An implementation and experimental artifacts published in a public repository, evaluated against general decomposition, single-entry single-exit (SESE) decomposition, and non-decomposed alignments on synthetic models, PLG-generated models, and BPI Challenge 2015 logs. ## Key findings - All decomposed approaches align faster than non-decomposed alignment on every synthetic dataset (models with 100 to 300 transitions), though no decomposed method is fastest everywhere. - The proposed decomposition yields fitness values closer to true non-decomposed fitness than SESE decomposition on both the pr[A-G]m6 models and the A/E 100-300 models; fewer subnets give closer approximations for both methods. - On the real-life BPI Challenge 2015 logs 1-4, SESE decomposition becomes slow because silent transitions in every fragment force merges, whereas the proposed subnets stay small. - With noisy logs and longer traces, the decomposed method slows down only slightly while non-decomposed alignment degrades substantially; limiting maximum subnet size speeds alignment but over-approximates fitness. - The ILP step scales poorly: decomposition takes about 2 s for 100 transitions, 9 s for 200, and 27 s for 300, exceeding SESE overhead, and validation on large real-world logs (around 300 activities) remains open. --- # PerSpect: An Interactive Tool for Performance Pattern Mining - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-caise-forum-perspect/ - Date: 2026-06-07 - Authors: Moritz Schlegelmilch, Piotr Zaniewski, Volodymyr Diemieniev, Jasper Saathoff, Tsung-Hao Huang, Gyunam Park - Venue: Intelligent Information Systems - CAiSE 2026 Forum and Doctoral Consortium, Verona, Italy, June 8-12, 2026, Proceedings - DOI: https://doi.org/10.1007/978-3-032-27997-2_18 - PDF: https://www.gyunam.com/publication/2026-caise-forum-perspect/paper.pdf - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) ## Abstract Complex processes can be difficult to understand from event data. Traditional process models effectively visualize control flows, but lack temporal context for performance analysis. To address this, the Performance Spectrum was introduced to map the dynamics of all steps and cases over time. However, its manual interpretation is challenging due to the overlapping patterns and noise. In this tool demonstration paper, we present PerSpect, an interactive, web-based user interface that significantly enhances the original Performance Spectrum approach. To facilitate interactive analysis, the tool automates the targeted detection of established operational behaviors (e.g., batching and FIFO), alongside advanced filtering capabilities and a side-by-side comparison function for evaluating the same segment across different process variants. ## Practical problem Process analysts who want to know why cases are slow usually look at a discovered process model, but such models show only the order of activities and nothing about time: they cannot reveal that cases pile up for months and are then released in one burst, or that work is handled strictly first-in-first-out in one period and chaotically in another. The Performance Spectrum visualization was invented to expose exactly these temporal patterns by drawing every case's hop between two activities on a timeline. In practice, however, the picture of a real log is a dense tangle of overlapping lines, so distinguishing batches from noise requires an expert staring at the screen, the existing ProM plugin and commercial implementation only highlight selections, and nothing lets the analyst extract the cases behind a pattern for further study. ## Research questions 1. How can the detection of operational patterns such as simultaneous, first-in-first-out, and non-first-in-first-out batching in a Performance Spectrum be automated instead of left to visual inspection? 2. How can analysts isolate, compare, and export the cases behind a temporal pattern across process variants within an interactive workflow? ## Contributions - `PerSpect`, an open-source web-based tool (GitHub) that reimplements the Performance Spectrum as the foundation of a pattern-oriented analysis workflow. - Automated batch detection in a spectrum segment via DBSCAN clustering on start, end, or both timestamps, with user-set epsilon and minimum-sample parameters and a FIFO versus non-FIFO distinction. - Global and segment-specific filters (variant, date range, duration quartile, interactive case selection) whose effects propagate across all displayed spectra, with tracked filter history. - Side-by-side comparison of named spectrum configurations, process and batch statistics, frequency diagrams, and export of the filtered cases as an XES event log. - A client-server architecture with a FastAPI backend, SQLite metadata store, in-memory LRU caching of logs, and a Vue single-page frontend. ## Key findings - This is a tool demonstration paper; it reports no controlled experiments, only a use case and a demonstration video. - On the Road Traffic Fine Management log, the batch filter on the segment Add Penalty to Send for Credit Collection shows that over 97% of cases follow a sequential batching pattern formed by only 11 batches with an average interval of 410 days, i.e. a long-term backlog cleared in rare bursts. - Setting epsilon to 0 minutes recovers simultaneous batches automatically, and overlapping start-batches and end-batches in the same segment can be separated and displayed individually. --- # Neuro-Symbolic Learning for Predictive Process Monitoring via Two-Stage Logic Tensor Networks with Rule Pruning - Type: Conference paper - URL: https://www.gyunam.com/publication/2026-pakdd-twostage/ - Date: 2026-06-07 - Authors: Fabrizio De Santis, Gyunam Park, Francesco Zanichelli - Venue: Advances in Knowledge Discovery and Data Mining - 30th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2026, Hong Kong, China, June 9-12, 2026, Proceedings, Part II - DOI: https://doi.org/10.1007/978-981-92-1462-4_9 - PDF: https://www.gyunam.com/publication/2026-pakdd-twostage/paper.pdf - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract Predictive modeling on sequential event data is critical for fraud detection and healthcare monitoring. Existing data-driven approaches learn correlations from historical data but fail to incorporate domain-specific sequential constraints and logical rules governing event relationships, limiting accuracy and regulatory compliance. For example, healthcare procedures must follow specific sequences, and financial transactions must adhere to compliance rules. We present a neuro-symbolic approach integrating domain knowledge as differentiable logical constraints using Logic Networks (LTNs). We formalize control-flow, temporal, and payload knowledge using Linear Temporal Logic and first-order logic. Our key contribution is a two-stage optimization strategy addressing LTNs’ tendency to satisfy logical formulas at the expense of predictive accuracy. The approach uses weighted axiom loss during pretraining to prioritize data learning, followed by rule pruning that retains only consistent, contributive axioms based on satisfaction dynamics. Evaluation on four real-world event logs shows that domain knowledge injection significantly improves predictive performance, with the two-stage optimization proving essential knowledge (without it, knowledge can severely degrade performance). The approach excels particularly in compliance-constrained scenarios with limited compliant training examples, achieving superior performance compared to purely data-driven baselines while ensuring adherence to domain constraints. ## Practical problem Banks, hospitals and manufacturers want to predict how a running case will end, for example whether a patient will develop complications or a loan will be rejected, from the event records their systems collect. Deep learning predictors learn correlations from past cases but know nothing about the mandatory rules that govern the process, such as a transfer requiring prior identity verification, and such rules are rarely violated in training data yet critical in live operation. Earlier attempts to inject knowledge handle only control-flow rules, add them only at the input or output, and cannot cope with noisy or conflicting rules. Simply adding logical rules to a neural model can even backfire, because the model satisfies the rules trivially instead of learning to predict. ## Research questions 1. Does injecting domain knowledge as logical constraints improve predictive performance compared to purely data-driven approaches? 2. What do the two-stage optimization and the rule pruning mechanism contribute to model performance? 3. How well does the approach generalize to compliance-constrained scenarios with few compliant training examples? ## Contributions - A scheme for extracting and categorizing process knowledge into control-flow, temporal and payload features and rules, formalized in Linear Temporal Logic (LTL) and first-order logic using Declare mining, SLA compliance analysis and statistical correlations. - A neuro-symbolic outcome classifier for predictive process monitoring built on Logic Tensor Networks (LTN) with LSTM or Transformer backbones, separating data axioms from knowledge axioms. - A two-stage optimization: weighted axiom loss (0.8 data, 0.2 knowledge) during pretraining, then rule pruning with an analytic gating score from mean and variance of rule satisfaction, followed by fine-tuning on the pruned knowledge base. - A compliance-aware evaluation protocol that enriches test sets with rule-compliant traces, used alongside a standard temporal split. - An open-source implementation on GitHub (FabrizioDeSantis/NeSy-PPM). ## Key findings - On a temporal split of Sepsis, BPIC2012, BPIC2017 and Traffic fines, Two-Stage variants beat LSTM and Transformer baselines on every log; F1 improves by 5.23% on Sepsis (LSTM to Two-Stage-L) and 10.26% on BPIC2012 (Transformer to Two-Stage-T). - Adding knowledge axioms without weighting or pruning (LTN-NoP) collapses performance to 18.79% and 19.18% F1 on Sepsis, below data-only LTN (69.10%), whereas Two-Stage-L reaches 75.82% F1 and Two-Stage-T 75.63%. - On the compliance-aware test set with only 4% compliant training traces (Sepsis), Two-Stage-L reaches 89.96% F1 versus 73.26% for LSTM and 37.52% for LTN-NoP-L; on BPIC2012 Two-Stage-T reaches 64.23% versus 55.61% for the Transformer. - LSTM backbones do better on short traces (Sepsis and Traffic fines, maximum length 13 and 10) and Transformers on longer traces (BPIC2012 and BPIC2017, maximum length 40 and 20). - Gains shrink on data-rich logs: on Traffic fines F1 improves by only 1.1% and 0.38%, so knowledge injection pays off mainly when training data or compliant examples are scarce. --- # Neuro-symbolic artificial intelligence for real-time sepsis mortality prediction - Type: Journal article - URL: https://www.gyunam.com/publication/2026-eaai-sepsis/ - Date: 2026-04-24 - Authors: Fabrizio De Santis, Gyunam Park, Francesco Zanichelli - Venue: Engineering Applications of Artificial Intelligence - DOI: https://doi.org/10.1016/j.engappai.2026.114920 - PDF: https://www.gyunam.com/publication/2026-eaai-sepsis/paper.pdf - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract Sepsis mortality prediction in intensive care units represents a critical engineering challenge in healthcare systems, where timely intervention through intelligent decision support systems can significantly reduce mortality rates. Current clinical decision support systems rely on black-box machine learning models that lack the interpretability required for safe deployment in life-critical engineering applications. We develop a neuro-symbolic artificial intelligence framework that integrates Logic Tensor Networks with automated clinical knowledge extraction for interpretable mortality prediction. Our novel contributions include: (1) an automated pipeline for transforming clinical guidelines into first-order logic constraints using medical knowledge graphs and rule mining; (2) a dual axiom architecture with weak anchoring that enables context-adaptive learning while maintaining semantic grounding; and (3) a three-tier explainability framework providing concept-level, rule-level, and patient-level interpretability. Using 19,328 sepsis patients from a publicly available real-world clinical dataset, our system achieves 86.45% accuracy at 6-h prediction windows, outperforming traditional machine learning baselines while providing full transparency in decision-making. External validation on 6735 patients demonstrates preserved ranking ability despite domain shift challenges. ## Practical problem Intensive care clinicians need to know early which sepsis patients are likely to die, so they can escalate treatment and allocate scarce resources in time. Bedside scoring systems such as SOFA and APACHE II use only a handful of variables and are only moderately accurate, while machine learning models are more accurate but behave as black boxes, so a physician cannot check the rationale behind a life-or-death prediction. Post-hoc explainers such as SHAP give feature importances rather than medical reasoning and can be misleading when features are correlated. Existing neuro-symbolic attempts rely on hand-written rules, ignore the irregular temporal dynamics of ICU data, and explain only at one level, which keeps them out of clinical use. ## Research questions 1. Can clinical guidelines be turned automatically into first-order logic constraints and embedded in a neural mortality predictor without manual rule engineering? 2. Does injecting such knowledge improve sepsis mortality prediction over purely data-driven baselines across prediction horizons, and which design choices make the integration work rather than harm? 3. Does the resulting model expose clinically verifiable reasoning at concept, rule and patient level, and does it transport to an external hospital cohort? ## Contributions - NeSy-SMP, a neuro-symbolic framework for sepsis mortality prediction that grounds Logic Tensor Networks (LTN) predicates in a bidirectional LSTM over temporal sequences of clinical events plus comorbidity features. - An automated knowledge pipeline that builds a SNOMED CT knowledge graph from the Surviving Sepsis Campaign guidelines, mines Horn rules with `AnyBURL` (confidence above 0.8) and translates them into first-order logic axioms. - A dual axiom architecture that separates data axioms from knowledge axioms with a weighted satisfiability loss (0.8/0.2) and weakly anchors each clinical predicate to an established threshold. - A three-tier explainability framework covering clinical concept learning verification, medical rule consistency validation and an individual patient reasoning dashboard. - An evaluation protocol over 6, 12, 24 and 48-hour lead times on MIMIC-IV with clinical operating-point analysis and external validation on eICU, with code released for reproduction. ## Key findings - On 19,328 MIMIC-IV sepsis patients with 5-fold cross-validation at threshold 0.5, NeSy-SMP has the highest F1 and recall at every lead time, reaching 86.45% accuracy and 80.35% F1 at 6 hours with AUC 88.33% comparable to XGBoost (88.56%). - Ablation shows knowledge axioms alone (LTN-K) lower performance below the base LTN, and weak anchoring with equal axiom weights (LTN-AK) collapses to always predicting death (19.03% F1, 50% recall); only anchoring plus weighting helps. - At clinical operating points for a 500-patient-per-year hospital (24-hour window), NeSy-SMP at 85% sensitivity produces 137 false alerts per year versus 190 for BiLSTM, and at 88% specificity catches 63 deaths versus 61 for XGBoost and Random Forest and 55 for BiLSTM. - Learned predicates align with clinical thresholds after training (lactate 4.0 mmol/L, bilirubin 2.0 mg/dL), and rule truth values separate deaths from survivors for lactate, bilirubin and white blood cells but not for C-reactive protein or mean arterial pressure. - On 6,735 external eICU patients without retraining, all models fall to AUC 62-72% and overestimate risk; tree ensembles keep the highest AUC, NeSy-SMP leads F1 only at 6 hours (61.26%), and rule activation drops sharply (chronic conditions 32.6% to 3.0%), while training costs 706 s versus 64 s for BiLSTM. --- # Configuring Large Reasoning Models Using Process Mining: A Benchmark and a Case Study - Type: Conference paper - URL: https://www.gyunam.com/publication/2025-bpm-workshop-lrm/ - Date: 2026-04-01 - Authors: Alessandro Berti, Humam Kourani, Gyunam Park, Wil M. P. van der Aalst - Venue: Business Process Management Workshops - DOI: https://doi.org/10.1007/978-3-032-13426-4_30 - PDF: https://www.gyunam.com/publication/2025-bpm-workshop-lrm/paper.pdf - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) ## Abstract Large Reasoning Models (LRMs), a subset of Large Language Models (LLMs) trained to articulate their chain-of-thought, have shown promise in tackling complex scientific tasks. However, evaluating and configuring their reasoning processes remains underexplored. This paper leverages a process mining-specific LLM evaluation framework to propose a methodology for analyzing and configuring LRMs. We introduce an approach to extract and classify reasoning steps by type (e.g., Deductive Reasoning, or Hypothesis Generation) and effect (Positive, Indifferent, Negative) on the overall reasoning, enabling a detailed assessment of reasoning quality. From this, we derive a new benchmark, PMLRM-Bench, which evaluates not only the correctness of outputs but also the robustness of the reasoning process. A case study on the QwQ-32B LLM demonstrates how targeted adjustments to reasoning type frequencies can boost task-specific performance. Our results reveal distinct reasoning patterns across models and provide actionable insights for LRM configuration. This work bridges process mining and LLM evaluation, offering a scalable framework for reasoning analysis. ## Practical problem Analysts increasingly ask large language models to read event logs, spot bottlenecks, check conformance or draft process models, and a newer class of these models, large reasoning models, writes out its chain of thought before answering. Existing benchmarks only grade the final answer, so a model that reaches the right conclusion through flawed or speculative steps scores the same as one that reasons soundly, and nobody can tell which reasoning habits a model should do more or less of. Teams who want to tune such a model for their domain therefore have no systematic way to diagnose its reasoning or to measure whether a configuration change helped. ## Research questions 1. How can the reasoning traces of large reasoning models be extracted and classified by reasoning type and by their effect on correctness so that reasoning quality can be measured, not just answer accuracy? 2. Do targeted adjustments to the frequency of specific reasoning types, applied through system prompts, change a model's performance on process mining tasks? ## Contributions - A reasoning analysis pipeline that parses chain-of-thought traces into steps and classifies each by one of ten reasoning types and by positive, indifferent or negative effect, using a judge LLM. - A composite score S that rewards correct conclusions and positive-effect steps while penalizing partially correct and wrong conclusions and indifferent or negative steps. - PMLRM-Bench, an extension of PM-LLM-Benchmark v2.0 covering 17 large reasoning models, released on GitHub (fit-alessandro-berti/pmllmbench-lrms-reasoning-analysis). - A case study protocol that varies the frequency of hypothesis generation, ethical reasoning and validation/backtracking in QwQ-32B via system prompts and re-evaluates the variants. - A validation of the judge's classifications with a second judge LLM. ## Key findings - Composite scores range from 14,459 (Grok-3-thinking) to -7,724 (R1-Distill-Qwen-1.5B), and they correlate strongly with PM-LLM-Benchmark scores (Pearson's r = 0.89), so robust reasoning tracks task performance. - Across models, deductive reasoning (25 to 40% of steps) and hypothesis generation dominate; weaker models produce far more negative-effect steps (438 for the 1.5B model versus 10 for the top model) and lean on speculative reasoning without validation. - A second judge (ChatGPT-4o) fully agrees with the Gemini-2.5-Pro classifications in 82.40% of steps, partially in 13.01%, and disagrees in 4.59%. - For QwQ-32B, prompting for more hypothesis generation raises the PM-LLM-Benchmark score from 36.9 to 37.1, while suppressing it lowers the score to 36.3 and the Hypotheses Generation category from 4.9 to 4.3. - Adjusting ethical reasoning does not help: less of it drops Fairness from 5.7 to 5.1, and more of it also lowers Fairness to 5.3 despite the highest composite score (14,218); results depend on the judge LLM's classification accuracy. --- # Extracting declarative constraints for process modeling from natural language descriptions with large language models - Type: Journal article - URL: https://www.gyunam.com/publication/2026-dss-declarativenl/ - Date: 2026-01-30 - Authors: Gyunam Park, Julian Kofferath, Minsu Cho - Venue: Decision Support Systems - DOI: https://doi.org/10.1016/j.dss.2026.114627 - PDF: https://www.gyunam.com/publication/2026-dss-declarativenl/paper.pdf - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md) ## Abstract With the growing availability of unstructured text data in organizations, automating the extraction of process models from natural language descriptions has become increasingly crucial. Traditional rule-based techniques face challenges such as limited generalization and constraints to narrow sets of control-flow constructs. We present a systematic empirical study of Large Language Models (LLMs) for translating natural language sentences into declarative process constraints using the Declare modeling language. Through comprehensive evaluation across seven LLM architectures, we demonstrate that fine-tuned models achieve 97.8% template accuracy compared to 53.8% for existing rule-based approaches. Our fine-tuning approach substantially outperforms prompting techniques (99.4% vs. 56.5% template accuracy), establishing clear guidance for practical deployment. We contribute a benchmark dataset of 969 sentence-constraint pairs across 11 Declare templates, direction-sensitive evaluation metrics, and complete reproducibility materials. The extracted constraints enable practical decision support applications including compliance monitoring, conformance checking, and automated governance dashboards. ## Practical problem Companies describe how their work should be done in text documents, handbooks and regulations, but analysis tools need formal process models, and the domain experts who know the processes rarely have the modeling skills or time to build them. Automated extraction tools based on hand-written language rules exist, yet they break on unfamiliar sentence structures or noisy wording and recognize only a handful of ordering constraints. There is also no shared dataset or metric for judging such tools, so it is hard to tell which one works. Analysts therefore still translate text into models manually before they can check compliance or compare documented and actual behavior. ## Research questions 1. Can a fine-tuned large language model translate natural language sentences into Declare constraints more accurately than rule-based pipelines and than prompting approaches? 2. Can such a model cover a broader range of Declare templates than existing extraction methods support? 3. How can constraints extracted sentence by sentence be merged into one coherent Declare process model with consistent activity labels? ## Contributions - A pipeline that translates each sentence of a process description into a Declare constraint with a fine-tuned decoder-only large language model (LLM), matches activity labels across constraints and merges them into a Declare process model. - Parameter-efficient fine-tuning with Low-Rank Adaptation (LoRA) of five open seven-to-eight-billion-parameter models (Gemma 7B, Mistral 7B, Falcon 7B, Llama-2 7B, Llama-3 8B) on a question-and-answer prompt format. - A publicly released benchmark dataset of 969 sentences paired with expert-annotated Declare constraints covering 11 templates across healthcare, compliance, education and finance. - An evaluation metric that weights template errors above activity errors and penalizes direction errors, which the slot-filling precision and recall used previously cannot detect. - Zero-shot and few-shot prompting strategies for matching semantically equivalent activity labels (verb-, noun- and gerund-based) across extracted constraints. ## Key findings - In 10-fold cross-validation on the 969-sentence benchmark, all five fine-tuned models reach template accuracy between 0.970 and 0.980; Gemma 7B scores highest with F1 0.884 (alpha 2), 0.918 (alpha 4) and template accuracy 0.980. - Against the rule-based state of the art of van der Aa et al. on validation set V1 (80 sentences; Init, End, Response, Precedence), fine-tuned Gemma 7B reaches template accuracy 0.988 versus 0.538; on V2 (45 sentences from their data) 0.956 versus 0.822. - Gains are largest for unary templates: on V1, Init improves from 0.273 to 0.910 and End from 0.0 to 1.0; the baseline ties on V2 Precedence (0.963) only because its fallback rule emits precedence constraints by default. - Fine-tuning beats prompting the same Gemma 7B over all 11 templates: template accuracy 0.994 versus 0.459 for few-shot and 0.565 for few-shot chain-of-thought prompting; F1 (alpha 2) 0.911 versus 0.268 and 0.598. - Limitation: the method assumes one constraint per sentence, so a sentence describing two constraints collapses into one; branched responses and time conditions are unsupported, and remaining template errors stem from linguistic ambiguity. --- # Reliable and Configurable Process Simulations via Probabilistic White-Box Models - Type: Conference paper - URL: https://www.gyunam.com/publication/2025-icsoc-whiteboxsim/ - Date: 2026-01-02 - Authors: Francesco Vinci, Gyunam Park, Wil M. P. van der Aalst, Massimiliano de Leoni - Venue: Service-Oriented Computing - DOI: https://doi.org/10.1007/978-981-95-5015-9_24 - PDF: https://www.gyunam.com/publication/2025-icsoc-whiteboxsim/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Business process simulation (BPS) has emerged as a crucial tool that offers a risk-free virtual environment to analyze, test, and optimize complex service compositions and orchestrations in cloud and edge computing environments. BPS enables the evaluation of alternative scenarios (a.k.a. “what-if” scenarios) by capturing the dynamic behavior of service interactions, including control-flow, task durations, and resource utilization. Several techniques to discover BPS models employ black-box predictors to characterize the run-time simulation aspects. The downside of these approaches is that the rules guiding the predictors cannot be provided in an intelligible form. This means that service architects and process analysts cannot modify or explore the decision-making logic that drives service orchestration and composition in these simulations. Moreover, these models are often deterministic and thus unable to capture uncertainty, which is essential to realistically simulate dynamic environments. This paper presents white-box predictors based on probabilistic decision trees, which are intelligible and easy to configure. The experiments show that white-box predictors can improve the simulation accuracy and variability, while being naturally intelligible. ## Practical problem Process analysts and service architects use simulation to test what-if scenarios such as adding staff, changing service-level agreements or automating a task without touching the live system. Recent simulators plug deep learning models into the simulation to predict durations and routing, and these are accurate but opaque: nobody can read or edit the rules, so the what-if scenarios the simulation exists for cannot be expressed. Such models also need large amounts of data and return the same output for the same input, so they fail to reproduce the natural variability of real operations. Classic white-box simulators are editable but rely on simple distributions and are noticeably less accurate. ## Research questions 1. Can probabilistic decision trees characterize the run-time behavior of a process simulation model (routing, execution times, waiting times, arrivals) while remaining interpretable and editable? 2. How does such a white-box simulation compare in accuracy with the white-box baseline Simod and with the black-box deep learning simulators DSim and RIMS? 3. Does probabilistic rather than deterministic prediction yield more realistic variability in the simulated behavior? ## Contributions - Decision-Aware Business Process Simulation (DBPS) models: a Petri net combined with probabilistic decision trees that set transition weights, activity execution times, resource-specific waiting times and inter-arrival times from trace history, attributes, resource, workload, weekday and hour. - A discovery technique that aligns the log to the control-flow model, trains CART decision trees per transition, activity and resource, and fits a per-leaf distribution (normal, exponential, log-normal, uniform or constant) by minimizing Wasserstein distance. - Three what-if interaction operations on the trees: changing leaf distributions, replacing a subtree by a leaf, and replacing a leaf by a subtree. - An open-source Python package (`prosit` on GitHub) that discovers a model from an event log and Petri net or lets users define each element manually. - An evaluation protocol on five logs (Purchase-to-Pay, a pharmacy process, academic credential recognition, and the BPI Challenge 2012 and 2017 workflow logs) with control-flow, cycle-time, arrival, rule-level and entropy measures. ## Key findings - Across the five logs, DBPS beats the white-box baseline Simod on all but two metric-log combinations; average N-gram distance (NGD) 0.28 versus 0.66, cycle time distribution distance (CTD) 153.83 versus 356.06, case arrival rate distance (CAR) 200.12 versus 232.37. - DBPS is competitive with the black-box simulators: average NGD 0.28 versus 0.49 (DSim) and 0.48 (RIMS), average CTD 153.83 versus 194.92 and 170.53, average CAR 200.12 versus 229.91 and 240.39. - At the level of decision rules, DBPS has lower average distances than Simod on all four rule metrics, most clearly control-flow rule distance 0.07 versus 0.23 and execution-time rule distance 9.70 versus 32.07. - Average execution-time entropy is 1.47 for DBPS versus 1.32 for Simod, 0.73 for DSim, 0.71 for RIMS and 0.57 for a deterministic DBPS variant, so the probabilistic leaves are what restores variability. - Limitation: on the synthetic pharmacy log DBPS has the lowest entropy (0.36 versus 0.64 for Simod), and the evaluation is purely quantitative; whether analysts can actually use trees of depth up to 5 for what-if analysis is untested. --- # GOProQ: A Graphical Query Language for Object-Centric Process Analysis - Type: Conference paper - URL: https://www.gyunam.com/publication/2025-icsoc-goproq/ - Date: 2026-01-02 - Authors: Gyunam Park, Jan Niklas Adams, Daniel Schuster - Venue: Service-Oriented Computing - 23rd International Conference, ICSOC 2025, Shenzhen, China, December 1-4, 2025, Proceedings, Part II - DOI: https://doi.org/10.1007/978-981-95-5015-9_21 - PDF: https://www.gyunam.com/publication/2025-icsoc-goproq/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Service-oriented architectures facilitate distributed business processes where multiple services collaborate to achieve business goals. However, traditional process mining approaches rely on single-case event logs that fail to capture the complex interactions between services and their associated data objects. Object-Centric Event Logs (OCELs) address this challenge by modeling service processes as graphs of interconnected events and objects, providing a foundation for comprehensive service analysis. While existing querying approaches offer capabilities for traditional event data, they lack the expressiveness needed for analyzing object-centric service interactions. This paper introduces GOProQ, a graphical query language specifically designed for querying object-centric service processes. We identify key requirements for service-oriented process querying, define a formal language that addresses these requirements, and implement an interactive tool for service analysts. Our evaluation across multiple real-world service datasets demonstrates GOProQ's expressiveness and performance for service monitoring in complex service-oriented environments. ## Practical problem Service analysts monitoring distributed processes need to ask questions that span several interacting objects at once, such as orders whose items come from multiple suppliers and whose payment arrived before all deliveries completed. Object-centric event logs (OCELs) record such executions as graphs of events connected through shared objects, but the available query languages were built for single-case logs or for general graph databases. Compliance rule graphs, DCR graphs and process-instance query languages assume one case per event, and general-purpose languages such as Cypher become hard to write and slow for multi-object patterns. Analysts therefore lack a way to express and run cross-service queries, for example for service-level agreement monitoring, without hand-written code. ## Research questions 1. Which querying dimensions and expressiveness requirements does a query language for object-centric service process executions have to cover? 2. How can a graphical query language be defined, with formal syntax and semantics, that covers nodes and edges labelled by both activities and object types, including cardinality, directly- and eventually-follows relations and logical composition? 3. Is such a language feasible in practice on real-world OCELs in terms of computation time, and how much does query reordering help? ## Contributions - A two-dimensional requirements framework for querying service process executions: structural elements (event nodes, precedence edges) crossed with label types (activities, object types), plus four expressiveness and three performance requirements. - GOProQ, a graphical query language with formal syntax and semantics over service process executions derived from OCELs, covering object-type queries, activity presence/start/end queries, directly-follows and eventually-follows queries with cardinality constraints, and AND/OR/NOT composition. - A query-reordering evaluation strategy that runs cheap object-type and activity queries before control-flow queries and NOT before AND before OR, to prune non-matching executions early. - An interactive tool to import OCELs, compose queries graphically, execute them and visualize matching executions, released with a user manual on GitHub (gyunamister/GOProQ). - An evaluation on six public OCELs (two loan-application logs, recruiting, travel reimbursement, procure-to-payment, order-to-cash; up to 392,018 events and 31,305 executions). ## Key findings - The tool covers every cell of the requirements matrix, including cardinality constraints, the directly- versus eventually-follows distinction and logical composition, so the expressiveness requirements are met by construction. - Object-type queries are the cheapest (0.0009 s to 0.2261 s per dataset), activity queries cost more (up to 12.6 s on the large loan log) and eventually-follows queries are the most expensive (20.7 s on the large loan log). - Eventually-follows queries could not be completed in reasonable time on the travel-reimbursement and procure-to-payment logs, whose executions contain thousands of nodes (8,481 nodes and 8,611 edges in one travel-reimbursement execution). - The reordering strategy lowers computation time on every dataset against random query order; on the recruiting log the median time drops by 57.1%, and with reordering all datasets except the large loan log answer queries in under 3 seconds. - Limitations: direct object-to-object relations not mediated by activities are not queryable, large executions still exceed interactive response times, and deeply nested graphical queries become cognitively demanding. --- # Prosit: A tool for interactive and transparent process simulations - Type: Conference paper - URL: https://www.gyunam.com/publication/2025-icsoc-prosit/ - Date: 2025-12-01 - Authors: Francesco Vinci, Gyunam Park, Wil M. P. van der Aalst, Massimiliano de Leoni - Venue: Proceedings of the 23rd International Conference on Service-Oriented Computing (ICSOC 2025), Demo Track. Lecture Notes in Computer Science, Springer, Cham - PDF: https://www.gyunam.com/publication/2025-icsoc-prosit/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Business Process Simulation (BPS) provides a controlled environment to evaluate alternative process designs and perform “what-if” analyses. While several approaches exist for deriving simulation models directly from event logs, existing solutions are often limited in accessibility, configurability, and interpretability. We present ProSiT: Process Simulation Tool, a user-friendly environment for data–driven simulation. ProSiT supports (i) the ingestion of event logs and process models, (ii) automated discovery of control-flow, timing, resource, and trace attribute parameters, (iii) interactive configuration of alternative service scenarios, (iv) visualization of accuracy metrics, and (v) generation and statistical analysis of simulated event logs. ProSiT integrates state-of-the-art machine learning methods for parameter discovery in an explainable whitebox form, contrasting with existing deep learning–based approaches that act as black boxes. By combining simulation accuracy with transparency, ProSiT empowers service architects and process analysts to experiment with alternative designs, evaluate service-level impacts, and explore optimization opportunities through an intuitive graphical interface. ## Practical problem Before changing how a service is staffed or how work is routed, analysts want to try the change in a simulation rather than on live operations. Building such simulation models by hand is slow and error-prone, so recent tools learn them from event logs, but the most accurate learners are deep neural networks whose decision logic nobody can read or edit. That defeats the purpose of what-if analysis: an architect who cannot see why the model routes a case to a particular resource cannot credibly modify that behavior to test an alternative. Analysts are thus left choosing between manually built models they understand and automatically discovered ones they cannot inspect, and most research prototypes are hard to install and configure in the first place. ## Research questions 1. How can data-driven discovery of simulation parameters be made transparent enough that analysts can inspect, validate, and modify every parameter before running what-if scenarios? 2. How can log ingestion, parameter discovery, scenario configuration, accuracy assessment, and simulation be offered as one accessible end-to-end workflow? ## Contributions - `ProSiT` (Process Simulation Tool), a web-based client-server application with a Flask backend, HTML/CSS/JavaScript frontend, SQLite persistence, and Docker distribution, with source code on GitHub. - A five-phase workflow: ingestion of XES logs and optional PNML models, parameter discovery, interactive scenario configuration, accuracy assessment, and simulation producing a downloadable synthetic log. - White-box parameter discovery based on probabilistic decision trees of controllable depth, covering transition weights, resource assignment, multitasking and calendars, arrival, execution and waiting time distributions, and trace attributes, with optional incremental discovery. - An accuracy assessment module comparing simulated and real logs by n-gram control-flow distance, temporal distribution metrics, resource handover realism, and entropy-based generalization. - A RESTful API (upload, discover, get and update parameters, simulate) enabling integration with external systems alongside the graphical interface. ## Key findings - This is a tool demonstration paper; no quantitative evaluation is reported, and the paper refers to companion work for the accuracy of the underlying white-box discovery method. - A demonstration video walks through the complete workflow on the BPI Challenge 2017 event log, from upload through parameter discovery, scenario editing, simulation, and result analysis. - The tool outputs process maps annotated with performance data, cycle time distributions, resource utilization heatmaps, and activity execution and waiting time analyses; user studies on usability and trust are left to future work. --- # Unlocking Non-Block-Structured Decisions: Inductive Mining with Choice Graphs - Type: Conference paper - URL: https://www.gyunam.com/publication/2025-bpm-powl/ - Date: 2025-08-31 - Authors: Humam Kourani, Gyunam Park, Wil M. P. van der Aalst - Venue: Business Process Management - DOI: https://doi.org/10.1007/978-3-032-02867-9_10 - PDF: https://www.gyunam.com/publication/2025-bpm-powl/paper.pdf - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) ## Abstract Process discovery aims to automatically derive process models from event logs, enabling organizations to analyze and improve their operational processes. Inductive mining algorithms, while prioritizing soundness and efficiency through hierarchical modeling languages, often impose a strict block-structured representation. This limits their ability to accurately capture the complexities of real-world processes. While recent advancements like the Partially Ordered Workflow Language (POWL) have addressed the block-structure limitation for concurrency, a significant gap remains in effectively modeling non-block-structured decision points. In this paper, we bridge this gap by proposing an extension of POWL to handle non-block-structured decisions through the introduction of choice graphs. Choice graphs offer a structured yet flexible approach to model complex decision logic within the hierarchical framework of POWL. We present an inductive mining discovery algorithm that uses our extension and preserves the quality guarantees of the inductive mining framework. Our experimental evaluation demonstrates that the discovered models, enriched with choice graphs, more precisely represent the complex decision-making behavior found in real-world processes, without compromising the high scalability inherent in inductive mining techniques. ## Practical problem Analysts who want to see how a process really runs let discovery algorithms draw a model from the event log, and inductive miners are popular because their models are always free of deadlocks and are produced quickly even for large logs. The price is that these miners can only express decisions as neatly nested blocks: every split must be closed by a matching join. Real processes rarely behave that way; an order may be cancelled on one branch while the other branch rejoins the main flow, and a block-structured model then has to allow impossible behavior such as cancelling and shipping the same order. Directly-follows graphs capture such branching naturally but become imprecise as soon as activities run in parallel or repeat, so practitioners currently choose between a model that is sound but too permissive and one that is faithful to branching but unsound and over-general. ## Research questions 1. How can non-block-structured decision points be represented within the hierarchical Partially Ordered Workflow Language (POWL) without losing soundness? 2. Can the Inductive Miner be extended to discover such decision structures while keeping its formal fitness guarantee? 3. Do the discovered models become more precise on real-life event logs, and does the extension affect the scalability of inductive mining? ## Contributions - Formalizes choice graphs, a directed start-to-end graph over sub-models that generalizes the exclusive-choice operator, and defines POWL 2.0 together with its language semantics. - A discussion of conversion of POWL 2.0 to workflow nets and of its expressiveness, covering sound state-machine and marked-graph workflow nets within the free-choice class. - A discovery algorithm that replaces the exclusive-choice cut of the POWL Inductive Miner with a valid choice graph cut computed from the directly-follows graph, restricted to acyclic graphs. - Proves, by induction over the recursion, that the extended miner preserves the Inductive Miner's fitness guarantee for every trace in the log. - An implementation in the Python library `powl` (pip-installable, Zenodo archived), a Streamlit web demo, and a published collection of all discovered models. ## Key findings - On 17 real-life event logs from data.4tu.nl, the choice-graph miner achieves equal or higher precision and F-score than the standard POWL Inductive Miner in every case; in 21 of 30 non-timeout cases the models are strictly more precise. - With directly-follows noise filtering at threshold 0.2, F-score gains are large on several logs, e.g. BPIC2012 from 0.41 to 0.62, BPIC2017 from 0.48 to 0.70, Hospital Billing from 0.71 to 0.95, Road Traffic Fine from 0.82 to 0.99. - On the Receipt Phase WABO log both fitness (0.81 to 0.88) and precision (0.25 to 0.55) improve; on BPIC2017 Offer Log F-score rises from 0.91 to 1.00 because offers can no longer be accepted before being sent and returned. - Runtime is unaffected without filtering and sometimes improves with filtering, e.g. BPIC2019 discovery drops from 287 seconds to 20 seconds. - Limitations: fitness occasionally drops slightly after filtering, conformance checking timed out on BPIC2018 and BPIC2019, discovery handles only acyclic choice graphs, and POWL 2.0 still forbids arbitrary interleaving of choice and concurrency splits and joins. --- # Online Discovery of Simulation Models for Evolving Business Processes - Type: Conference paper - URL: https://www.gyunam.com/publication/2025-bpm-sim/ - Date: 2025-08-31 - Authors: Francesco Vinci, Gyunam Park, Wil M. P. van der Aalst, Massimiliano de Leoni - Venue: Business Process Management - DOI: https://doi.org/10.1007/978-3-032-02867-9_27 - PDF: https://www.gyunam.com/publication/2025-bpm-sim/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Business Process Simulation (BPS) refers to techniques designed to replicate the dynamic behavior of a business process. Many approaches have been proposed to automatically discover simulation models from historical event logs, reducing the cost and time to manually design them. However, in dynamic business environments, organizations continuously refine their processes to enhance efficiency, reduce costs, and improve customer satisfaction. Existing techniques to process simulation discovery lack adaptability to real-time operational changes. In this paper, we propose a streaming process simulation discovery technique that integrates Incremental Process Discovery with Online Machine Learning methods. This technique prioritizes recent data while preserving historical information, ensuring adaptation to evolving process dynamics. Experiments conducted on four different event logs demonstrate the importance in simulation of giving more weight to recent data while retaining historical knowledge. Our technique not only produces more stable simulations but also exhibits robustness in handling concept drift, as highlighted in one of the use cases. ## Practical problem Companies use business process simulation to test what-if scenarios, anticipate bottlenecks, and plan staffing, and tools now build those simulation models automatically from historical event data. The catch is that processes keep changing: a bank automates more reviews, a hospital reorganizes a department, workloads shift. A simulation model discovered from all past data blends the old and the new behaviour and then forecasts a mixture that never existed, while a model built only from the most recent weeks throws away hard-won knowledge about resources, calendars, and rare paths. Analysts are left either re-discovering models from scratch at every change or trusting models that drift out of date. ## Research questions 1. How can a hybrid business process simulation model be updated continuously from an event stream so that it adapts to recent behaviour while retaining historical knowledge? 2. Does prioritizing recent data while preserving historical information produce more accurate and more stable simulations than discovering from all past data or from only the latest window? ## Contributions - Formalizes online simulation-model discovery as an update function over a sliding window of an event stream, producing a sequence of models each extending the previous one. - A four-step update technique: trace-fragment extraction, incremental control-flow discovery (Schuster et al.), descriptive-parameter refresh, and online updating of predictive parameters. - Uses Hoeffding Adaptive Trees with adaptive windowing (ADWIN) drift detection as explainable, incrementally trained models for durations, waiting times, arrivals, and branching probabilities. - An implementation in Python on top of the `Cortado` library and the `River` online-learning library, released on GitHub (`franvinci/ProcessSimulationTool`). - An evaluation protocol with ten temporal windows and eight simulation-quality distances (Chapela-Campa et al.) comparing single-batch, last-batch, and online discovery on four event logs. ## Key findings - Across BPIC17W, BPIC12W, Purchase-to-Pay, and CVS pharmacy logs, the online technique is best in 24 of 32 log-metric combinations (75%), versus 8 for single-batch and 4 for last-batch discovery. - Temporal perspectives benefit most: average Cycle Time Distribution distance drops to 196.49 (online) from 245.87 (single batch) and 319.98 (last batch), and Relative Event Distribution to 156.16 from 195.77 and 241.06. - Control-flow distances stay close to single-batch results (average CFLD 0.14 vs 0.16; 3GD 0.17 vs 0.17), indicating no major control-flow drift in these logs. - Using only the last window is consistently worst (average CFLD 0.26, AED 2139.58 with CVS at 5711.09), showing that discarding historical data harms simulation quality. - On BPIC17W, after the Validate Application duration drops 38% at week 28, the online model yields lower Wasserstein distances than both baselines in all but two later weeks; removing obsolete behaviour from the incrementally updated model remains open. --- # Translating Workflow Nets into the Partially Ordered Workflow Language - Type: Conference paper - URL: https://www.gyunam.com/publication/2025-petri/ - Date: 2025-06-01 - Authors: Humam Kourani, Gyunam Park, Wil M. P. van der Aalst - Venue: Application and Theory of Petri Nets and Concurrency - 46th International Conference, PETRI NETS 2025, Paris, France, June 22-27, 2025, Proceedings - DOI: https://doi.org/10.1007/978-3-031-94634-9_12 - PDF: https://www.gyunam.com/publication/2025-petri/paper.pdf - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) ## Abstract Process discovery aims to automatically derive process models from event logs, enabling organizations to analyze and improve their operational processes. Inductive mining algorithms, while prioritizing soundness and efficiency through hierarchical modeling languages, often impose a strict block-structured representation. This limits their ability to accurately capture the complexities of real-world processes. While recent advancements like the Partially Ordered Workflow Language (POWL) have addressed the block-structure limitation for concurrency, a significant gap remains in effectively modeling non-block-structured decision points. In this paper, we bridge this gap by proposing an extension of POWL to handle non-block-structured decisions through the introduction of choice graphs. Choice graphs offer a structured yet flexible approach to model complex decision logic within the hierarchical framework of POWL. We present an inductive mining discovery algorithm that uses our extension and preserves the quality guarantees of the inductive mining framework. Our experimental evaluation demonstrates that the discovered models, enriched with choice graphs, more precisely represent the complex decision-making behavior found in real-world processes, without compromising the high scalability inherent in inductive mining techniques. ## Practical problem Organizations document and analyze their processes in established notations such as Petri nets (specifically workflow nets) and the Business Process Model and Notation (BPMN), and the tools they rely on expect those formats. These notations, however, give no guarantee that a model is free of deadlocks or other anomalies, and their flat graph structure makes large models hard for people to read and for algorithms to exploit. The Partially Ordered Workflow Language (POWL) is a hierarchical notation that fixes both issues and already powers process discovery and modeling from text, but adopting it would mean replacing existing models and tools. Turning a POWL model into a workflow net is easy; going the other way, so that existing models can be reused, had no algorithm with formal guarantees. ## Research questions 1. How can a safe and sound workflow net be decomposed and translated into a POWL model whose language equals that of the net? 2. For which class of workflow nets does such a translation always succeed, and how does the approach scale to large models compared with existing workflow-net-to-process-tree conversion? ## Contributions - A recursive top-down algorithm that identifies exclusive-choice, loop and partial-order patterns in a safe and sound workflow net and projects the net onto sub-nets for further decomposition. - Formal definitions of the three patterns and their projections, plus reduction rules (explicit decision-point places, self-loop handling) that widen the set of nets the algorithm accepts. - Proofs that each projection preserves safeness and soundness and that each pattern preserves language, yielding a correctness theorem for every produced POWL model. - Defines the class of semi-block-structured workflow nets, which contains equivalents of all POWL models, and proves the algorithm is complete on that class. - An open-source implementation (GitHub humam-kourani/WF-net-to-POWL), also integrated into the `ProMoAI` tool, evaluated against the `PM4Py`-based workflow-net-to-process-tree converter. ## Key findings - Every POWL model produced by the algorithm has exactly the language of the input workflow net (Theorem 1, by structural induction over the XOR, loop and partial-order cases). - The algorithm converts every semi-block-structured workflow net, a class that strictly contains block-structured nets and sound marked-graph workflow nets and sits inside sound free-choice nets (Theorem 2). - On 1000 generated workflow nets with 21 to 370 transitions and 15 to 305 places, all are converted in 0.002 to 2.48 seconds, versus 0.013 to 126 seconds for the process-tree converter. - On the 20 ground-truth workflow nets of a process-modeling benchmark, the algorithm converts all 20, whereas the process-tree converter handles only 17. - The algorithm returns null for nets whose language has no POWL equivalent, e.g. free-choice nets with non-block-structured decision points or long-term dependencies between choices. --- # Operational process monitoring: An object-centric approach - Type: Journal article - URL: https://www.gyunam.com/publication/2024-comind/ - Date: 2025-01-01 - Authors: Gyunam Park, Wil M. P. van der Aalst - Venue: Computers in Industry - DOI: https://doi.org/10.1016/j.compind.2024.104170 - PDF: https://www.gyunam.com/publication/2024-comind/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md), [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) - Tags: Operational process monitoring, Object-centric process mining, Decision support systems, Operational support ## Abstract In business processes, an operational problem refers to a deviation and an inefficiency that prohibits an organization from reaching its goals, e.g., a delay in approving a purchase order in a Procure-To-Pay (P2P) process. Operational process monitoring aims to assess the occurrence of such operational problems by analyzing event data that record the execution of business processes. Once the problems are detected, organizations can act upon the corresponding problems with viable actions, e.g., adding more resources, bypassing problematic activities, etc. A plethora of approaches have been proposed to implement operational process monitoring. The lion’s share of existing approaches assumes that a single case notion (e.g., a purchase order in a P2P process) exists in a business process and analyzes operational problems defined over the single case notion. However, most real-life business processes manifest the interplay of multiple interrelated objects. For instance, an execution of an omnipresent P2P process involves multiple objects of different types, e.g., purchase orders, goods receipts, invoices, etc. Applying the existing approaches in these object-centric business processes results in inaccurate or misleading results. In this study, we propose a novel approach to assessing operational problems within object-centric business processes. Our approach not only ensures an accurate assessment of existing problems but also facilitates the analysis of object-centric problems that consider the interaction among different objects. We evaluate this approach by applying it to both simulated business processes and real-life business processes. ## Practical problem Process owners want to be told automatically when something goes wrong in their operations: a purchase order approved too slowly, an invoice cleared against several unrelated invoices, goods received without a matching order. Existing monitoring tools let analysts define such problems, but they assume every event belongs to exactly one case, while real processes such as purchase-to-pay interleave purchase orders, goods receipts, invoices and materials in the same events. Forcing this reality into a single case notion drops events, duplicates them or fabricates causal links, so the monitor reports problems that never happened and misses ones that did. Problems about how many and which objects take part in an activity cannot even be expressed in these tools. ## Research questions 1. Which operational problems arise in object-centric business processes, including problems about the involvement and cardinality of objects in activities? 2. How can such problems be represented formally yet graphically, and evaluated directly on object-centric event logs without flattening to a single case notion? 3. How accurate is object-centric monitoring compared with single-case approaches, and how sensitive are its results to threshold values and monitoring window length? ## Contributions - A taxonomy of object-centric operational problems covering occurrence, ordering relation, object involvement (absence, presence, cardinality) and time- and frequency-related performance problems. - A set of object-centric behavioral metrics over Object-Centric Event Logs (OCELs), including flow, sojourn, synchronization, pooling, lagging and readiness time derived from enabling events. - Operational Problem Graphs (OPGs), a hypergraph notation with activity and object type nodes and three edge types, with formal occurrence semantics and a monitoring function over OCELs. - A web application (`ProAct` monitoring engine) with an OPG editor, repository and assessment and visualization module, released with documentation, source code and the simulated and real-life OCELs. - Use cases on a simulated manufacturing process and a simulated purchase-to-pay process on SAP ERP (enterprise resource planning) plus a real-life loan application process. ## Key findings - On the simulated manufacturing process (four object types, three monthly logs) and purchase-to-pay process (five object types, three logs), monitoring matches ground truth for all four and all three defined problems in every time window. - On the BPI Challenge 2017 loan process (Jan. 2016 to Dec. 2017), a single-case Declare baseline misses 'Assess Potential Fraud precedes Cancel Application' (metric 0.0001) while the object-centric approach detects it (0.6690). - The single-case baseline reports a sojourn time of 25 days 13:58:03 for Cancel Application against 24 days 23:09:35 computed object-centrically; five of the eight problems cannot be evaluated by the baseline at all. - Detection is highly threshold-sensitive: across 54 weekly logs, occurrences of 'Existence of Cancel Application' drop sharply between thresholds 0.3 and 0.4, and 'Assess Potential Fraud precedes Cancel Application' between 0.6 and 0.7. - Extending the monitoring window from 1 to 5 weeks does not raise occurrence frequency proportionally; some problems fluctuate or decline, and the OPG notation cannot yet express problems that span several object types or use the object-to-object qualifiers of the newer OCEL standard. --- # Learning recommendations from educational event data in higher education - Type: Journal article - URL: https://www.gyunam.com/publication/2024-jiis-learningrecommendation/ - Date: 2024-08-01 - Authors: Gyunam Park, Lukas Liss, Wil M. P. van der Aalst - Venue: Journal of Intelligent Information Systems - DOI: https://doi.org/10.1007/s10844-024-00873-w - PDF: https://www.gyunam.com/publication/2024-jiis-learningrecommendation/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract This paper presents a novel approach for generating actionable recommendations from educational event data collected by Campus Management Systems (CMS) to enhance study planning in higher education. The approach unfolds in three phases: feature identification tailored to the educational context, predictive modeling employing the RuleFit algorithm, and extracting actionable recommendations. We utilize diverse features, encompassing academic histories and course sequences, to capture the multi-dimensional nature of student academic behaviors. The effectiveness of our approach is empirically validated using data from the computer science bachelor's program at RWTH Aachen University, with the goal of predicting overall GPA and formulating recommendations to enhance academic performance. Our contributions lie in the novel adaptation of behavioral features for the educational domain and the strategic use of the RuleFit algorithm for both predictive modeling and the generation of practical recommendations, offering a data-driven foundation for informed study planning and academic decision-making. ## Practical problem University staff who design degree programs must decide which courses to recommend in which semester, how heavy a course load to allow, and which prerequisites to enforce, yet they usually do this from regulations and intuition rather than from evidence about what actually helped past students. Campus management systems record every exam attempt of every student, but existing analyses of such data look at single courses or grades in isolation and miss the order in which courses are taken and how they interact. The one prior attempt to mine course sequences relied on decision trees, which overfit and react strongly to unusual students, so the rules they produce are hard to trust and act on. Program designers therefore lack reliable, readable rules linking study behavior to outcomes such as grade point average. ## Research questions 1. Which behavioral features derived from campus management event data capture students' academic histories, course sequencing, and attempt patterns in a way that is useful for study planning? 2. Can an interpretable rule-ensemble model predict academic outcomes such as grade point average more accurately and robustly than the decision-tree approach used so far? 3. How can the rules of such a model be turned into actionable recommendations for study program designers? ## Contributions - A three-phase approach that defines objectives, engineers behavioral features from educational event logs, and distills recommendations from an interpretable predictive model. - A catalogue of educational behavioral features adapted from a general process mining framework, spanning control-flow (occurrences, first and last occurrence, directly-follows, eventually-follows), data (credits, grades), and time (study and course duration) perspectives. - Uses the RuleFit algorithm (gradient-boosted rules plus a Lasso-regularized sparse linear model) for predicting grade point average and ranking rules by importance and coefficient. - A recommendation generation procedure that interprets high-importance rules by coefficient sign and magnitude into guidance on course load and course sequencing. - An empirical study on campus management data of 3,190 computer science bachelor students at RWTH Aachen University, with per-semester datasets and five-fold cross-validation. ## Key findings - On five semester-stage datasets (Data-S2 to Data-S6) predicting grade point average up to semester N, RuleFit yields lower mean squared error than the decision-tree baseline in every case under five-fold cross-validation. - The gap is largest on Data-S3, where the decision tree's mean squared error exceeds 2 while RuleFit stays well below; variance of results is higher on Data-S2 and Data-S5, attributed to immature behavior patterns and a smaller cohort. - The number of extracted rules, each a candidate recommendation, ranges from 65 for the third semester to 143 for the fifth; most semesters produce more positive than negative rules, the fifth semester being the exception. - For fourth-semester students, the top rule (completing SAP by the second semester with more than 77 total credits) has importance 0.0898 and coefficient -0.1799, i.e. it is associated with a better grade point average. - Limitations: the evaluation covers a single program at one university, rules learned from history can become obsolete as curricula change, and personal, socio-economic, and instructor factors are not captured in the event data. --- # Super Variants - Type: Conference paper - URL: https://www.gyunam.com/publication/2024-bpm/ - Date: 2024-06-06 - Authors: Jan Niklas Adams, Emilie Hastrup-Kiil, Gyunam Park, Wil M. P. van der Aalst - Venue: 22nd Business Process Management Conference 2024, BPM2024, Krakow, Poland, 1-6 September, 2024, Proceedings - DOI: https://doi.org/10.1007/978-3-031-05760-1_10 - PDF: https://www.gyunam.com/publication/2024-bpm/paper.pdf - Link (URL): https://doi.org/10.1007/978-3-031-05760-1_10 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Process mining offers methods to analyze the actual controlflow behavior of a process. The two main available methods are process discovery and variant analysis. While process discovery aggregates all variants into one model, the models often suffer from high complexity and lack detailed granularity. Conversely, variants are simple and offer fine granularity but their sheer number makes them difficult to comprehend. To bridge this gap, we introduce the concept of super variants. These represent a middle ground between the complexity of discovered process models and the simplicity of variants, offering an aggregation of closely related variants. We propose a super-variant mining framework based on object-centric variants, evaluate its scalability, and demonstrate its utility through a practical use case. This new approach promises to enhance control-flow analysis by striking a new balance between complexity and aggregation. ## Practical problem Process analysts who want to understand how a process really runs have two options, and both fail at scale. A discovered process model folds every observed behavior into one diagram, which quickly becomes too complex to read or too loose to be informative. A variant list shows each distinct sequence of activities plainly, but real logs contain hundreds or thousands of variants, far more than a person can take in. Trace clustering groups similar behavior top-down, yet the clusters cannot be traced back to the individual variants, so an analyst cannot drill from a cluster with poor performance down to the concrete behavior that causes it. ## Research questions 1. How can closely related variants be aggregated into a representation that is more compact than a variant list but simpler than a discovered process model? 2. How can such aggregations be mined automatically from object-centric variants and arranged in a hierarchy that connects behavioral clusters to individual variants? 3. Is super variant mining computationally feasible on real event logs, and does it yield meaningful aggregation on a real process? ## Contributions - The concept of super variants: object-centric variants extended with choice (including optionality) and object multiplicity through super objects, with an equivalence notion reducible to labeled graph isomorphism. - A super variant mining framework with intra-variant summarization (generating potential super variants and a hitting-set selection) and inter-variant summarization (a Levenshtein-based distance, a matching problem with a non-assignment penalty, recursive merging into a hierarchy). - A visualization that extends the object-centric variant lane notation with star-marked super objects, vertically split chevrons for choices and dotted borders for optional activities. - An open-source implementation on top of the `ocpa` library, released on GitHub (EmilieHK/Object-Centric-Super-Variants). - A proof of concept on a public procure-to-pay OCEL 2.0 event log and a scalability evaluation that decomposes computation time by framework step. ## Key findings - In the running order-to-cash example, ten traditional variants reduce to four object-centric variants, then to two super variants and finally one super variant with an arbitrary number of items and an optional clearance activity. - On the procure-to-pay log (516 process executions, 314 object-centric variants, 5,128 objects of 7 types), four variants merge over three hierarchy levels into one super variant covering 20% of the control-flow behavior, adding only a material super object and one choice between approving and delegating a purchase order. - In the scalability study (five hierarchies of five levels from 32 random variants), intra-variant summarization time is dominated by generating potential super variants and grows with the number of generated summarizations and of events per variant. - Inter-variant distance calculation and matching cost rise with the number of super variants per level, and at the final level inter-variant summarization dominates all other steps, making it the scalability bottleneck for large logs. - Limitations: super variants with multiplicity leave open whether objects are batched, and the added value for end users has not yet been evaluated. --- # Conformance Checking and Performance Analysis Using Object-Centric Directly-Follows Graphs - Type: Conference paper - URL: https://www.gyunam.com/publication/2024-bpm-forum/ - Date: 2024-06-06 - Authors: Gyunam Park, Jan Niklas Adams, Wil M. P. van der Aalst - Venue: 22nd Business Process Management Forum 2024, BPM Forum2024, Krakow, Poland, 1-6 September, 2024, Proceedings - DOI: https://doi.org/10.1007/978-3-031-70418-5_11 - PDF: https://www.gyunam.com/publication/2024-bpm-forum/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md), [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) ## Abstract Traditional process mining often simplifies the multi-object reality of business processes, linking each event to a single object called “the case”. This oversimplification may lead to inaccurate analyses, caused by missing interactions between object types. In contrast, object-centric process mining permits events to be tied to multiple objects, capturing complex interactions and providing a more accurate representation of business processes. This paper introduces an approach for supporting object-centric process mining utilizing Object-Centric Directly-Follows Graphs (OC-DFGs). Despite their advantages, e.g., simplicity, OC-DFGs have been relatively untapped for essential process mining tasks, such as conformance checking and performance analysis. In order to address this, our research presents a comprehensive approach for OC-DFG-based conformance checking and performance analysis. We fully implement the proposed approach as a web application and demonstrate the use of OC-DFGs for these tasks within a case study of a real-life loan application process. ## Practical problem Process analysts in commercial tools work mostly with directly-follows graphs: simple process maps that show which activity tends to follow which, and that can be filtered interactively even on logs with millions of events. Real processes, however, involve several kinds of business objects at once, such as a loan application and the multiple offers attached to it, and forcing each event onto a single case duplicates events and hides interactions. Object-centric directly-follows graphs keep the simplicity while showing each object type as its own layer, but until now they could only be discovered, not used to check where reality deviates from the model or to locate delays. Analysts who wanted those answers had to switch to heavier formalisms such as object-centric Petri nets, which are harder to read and configure. ## Research questions 1. How can conformance checking based on alignments be performed and visualized on object-centric directly-follows graphs? 2. How can those alignments be used to compute object-centric performance measures such as synchronization, pooling and lagging time in the context of the model? ## Contributions - A conformance checking approach that projects an object-centric directly-follows graph per object type, translates each projection into a workflow net, aligns the projected event log, and maps log and model moves back onto the graph. - A visualization scheme distinguishing five kinds of log moves (dotted edges, self-loops) and model moves (double-line edges) on the object-centric graph. - A performance analysis approach that projects alignments onto the model, builds per-event relationships to preceding events, and computes sojourn, synchronization, flow, lagging and pooling time. - `Explori`, an open-source web application on top of the `ocpa` Python library (GitHub gyunamister/Explori) with log management, discovery, conformance checking and performance analysis. - A case study applying the approach to the BPI Challenge 2017 loan application process with applications and offers as object types. ## Key findings - On a loan application log of 31,010 events, 2,302 applications and 5,128 offers, the alignments on the object-centric directly-follows graph reveal 548 applications skipping submit application and 325 applications needing an extra handle leads activity. - For offers, 274 are cancelled immediately after creation, bypassing send offer and call after offers; 261 offers show an extra return offer and 107 applications an extra call incomplete files after cancellation. - The average sojourn time before cancel application is about 26 days and 23 hours, with a synchronization time of 5 days and 7 hours across the three preceding activities. - Offers pool for 5 days and 6 hours on average while the lagging time of offers is only 4 minutes versus 5 days and 6 hours for applications, showing that applications are cancelled only after all offers are consolidated. - The decompose-and-recombine design can yield incoherent global alignments when object types disagree on a shared activity, and the sequential workflow-net translation misrepresents concurrency in performance figures. --- # Unlocking learner engagement and performance - A multidimensional approach to mapping learners to learning cohorts - Type: Journal article - URL: https://www.gyunam.com/publication/2024-kim-unlocking/ - Date: 2024-06-01 - Authors: Hyomin Kim, Gyunam Park, Minsu Cho - Venue: Education and Information Technologies - DOI: https://doi.org/10.1007/s10639-024-12782-0 - PDF: https://www.gyunam.com/publication/2024-kim-unlocking/paper.pdf - Link (URL): https://link.springer.com/article/10.1007/s10639-024-12782-0 - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) - Tags: Education ## Abstract Learning analytics, located at the intersection of learning science, data science, and computer science, aims to leverage educational data to enhance teaching and learning. However, as educational data increases, distilling meaningful insights presents challenges, particularly concerning individual learner differences. This work introduces a comprehensive approach for designing and automatically mapping learners into meaningful cohorts based on diverse learning behaviors. It defines four critical contexts, including engagement, direction, repetitiveness, and orderliness, and generates practical learning cohorts from their combinations. The approach employs a time-series-based clustering method with K-means clustering using dynamic time warping to identify similar learning patterns. Statistical techniques like the Mann-Kendall test and Theil-Sen estimator further refine the process. A case study on data science courses validates the approach, offering novel insights into learner behavior. The contributions include a novel time series approach to characterizing learning behavior, new learning cohorts based on critical contexts, and a systematic method for automated cohort identification. ## Practical problem Instructors and online-education platforms collect huge volumes of clickstream and video-watching data, but turning it into something actionable means dealing with the fact that learners behave very differently from one another. The usual answer is cohort analysis: grouping learners with similar behaviour so that interventions, course design, and support can be targeted. In practice, however, cohorts are defined almost entirely by how much a learner engaged (samplers, completers, dropouts) and are often identified by hand. Other aspects of behaviour that teachers care about, such as whether learners follow the intended lecture order, revisit material, or jump around erratically, are left out, and the resulting groups are too coarse to explain why some learners succeed and others leave. ## Research questions 1. How effectively does a time-series clustering method categorize students into pre-defined, multidimensional learning cohorts in a real-world online course setting? 2. What insights into learner behaviour, achievement, and interaction can be derived from analyzing the identified learning cohorts? ## Contributions - A taxonomy of nine learning cohorts (for example Orderly Engagers, Recurrent Disorderly Observers, Contrary Learners) derived from four behavioural dimensions: engagement, direction, repetitiveness, and orderliness. - A representation of learning behaviour as a time series over logical time, mapping the sequence of lectures a learner watched against the course's prescribed order. - A method that segments learners by record length, clusters them with K-means under dynamic time warping, and computes DTW barycenters per cluster. - A rule-based mapping of barycenters to cohorts using the Mann-Kendall trend test, the Theil-Sen slope estimator, and a log-difference fluctuation measure. - A case study on three data-science courses (18,405 learners, 171,440 video-watching events) from a real online education platform, including process models discovered with Fluxicon `Disco`. ## Key findings - The method assigns learners of all three courses to the nine cohorts; Recurrent Orderly Observers are the largest cohort everywhere (16.6%, 17.1%, and 14.8% of learners in courses A, B, and C). - Over half of the learners in every course (51.6%, 62.9%, 58.5%) watch under 10% of the material and cannot be cohorted, highlighting early attrition on the platform. - Of 670 learners enrolled in two or more courses, only 162 (26%) keep the same cohort; a chi-squared test (178.68 vs. critical 6.63, p < 0.01) shows cohorts depend more on the course than on the learner. - Recurrent Orderly Engagers obtain the highest quiz scores and pass ratios in courses A and C and the highest interaction scores in all three courses; in project-based course B, Recurrent Disorderly Engagers score highest. - Cluster counts and thresholds (for example the 0.6 log-difference cut-off) still need expert judgement, dynamic time warping is costly at scale, no ground-truth cohorts exist for validation, and cohorts are static over a course. --- # Improving Predictive Process Monitoring Using Object-Centric Process Mining - Type: Conference paper - URL: https://www.gyunam.com/publication/2024-ecis/ - Date: 2024-06-01 - Authors: Jan Niklas Adams, Gyunam Park, Wil M. P. van der Aalst - Venue: The 32nd European Conference on Information Systems, ECIS2024, Paphos, Cyprus, 13-19 June, 2024, Proceedings - Link (URL): https://aisel.aisnet.org/ecis2024/track08_bpm_di/track08_bpm_di/7/ - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md), [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) --- # Overstock Problems in a Purchase-to-Pay Process: An Object-CentricProcess Mining Case Study - Type: Conference paper - URL: https://www.gyunam.com/publication/2024-caise-w/ - Date: 2024-01-01 - Authors: Dina Kretzschmann, Gyunam Park, Alessandro Berti, Wil M. P. van der Aalst - Venue: Advanced Information Systems Engineering Workshops - CAiSE 2024 InternationalWorkshops, Limassol, Cyprus, June 3-7, 2024, Proceedings - DOI: https://doi.org/10.1007/978-3-031-61003-5_29 - PDF: https://www.gyunam.com/publication/2024-caise-w/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract This paper addresses overstock issues in a real-life Purchase-to-Pay process in cooperation with the industry leader in pet retail in Europe. It highlights the development of solutions for more efficient inventory management, thereby reducing overstock. Our approach involves identifying patterns leading to overstock and proposing specific improvement measures within the existing logistics systems. This includes technical modifications in the order suggestion and purchase order processes using Logomate and SAP systems. The research utilizes object-centric process mining techniques as a crucial tool to uncover these patterns, with a focus on the practical solutions derived for overstock reduction. The case study conducted with the PM2 methodology demonstrates potential benefits in optimizing inventory structure and suggests a path for future research in generalizing these findings across various sectors and automating overstock pattern detection. ## Practical problem Retailers keep warehouses full to avoid empty shelves, and the result is often overstock worth millions of euros: goods bought without matching demand that tie up capital and reduce flexibility. Logistics managers know the problem exists but struggle to see which concrete ordering behaviors create it, because purchasing runs across several systems (a replenishment planner proposing orders, an ERP system creating and receiving them) and a single purchase order bundles many items from many suppliers in many shipments. Classical inventory formulas assume constant demand and costs, and process mapping or fishbone diagrams are not driven by data. Conventional process mining looks at one order at a time and cannot follow a material that is re-ordered across several purchase orders, so the chains of events that end in overstock stay invisible. ## Research questions 1. How do the relevant objects in the purchase-to-pay process (order suggestions, order suggestion items, purchase orders, purchase order items, materials) behave and interact with each other? 2. Which behaviors and interactions of these objects lead to overstock? ## Contributions - A real-life object-centric process mining case study at Europe's largest pet retailer, following the PM2 project methodology adapted to the retailer's setting. - A conceptual model of the purchase-to-pay process with five object types and their object-to-object and event-to-object cardinalities, covering both the order suggestion and purchase order sub-processes. - An extraction and transformation pipeline from `Celonis` (fed by Logomate and SAP) into an object-centric event log in the OCEL standard format. - Two object-centric Petri nets, discovered and analyzed with the `ocpa` Python library, giving a holistic view of ordering and a material-focused view of receipt behavior. - A catalogue of assessed behavior patterns in both sub-processes and two concrete improvement measures agreed with logistics management for the Logomate and SAP systems. ## Key findings - The filtered log (January to April 2023, one central warehouse, short lead-time standard assortment) contains 13,151 events over 1,918 order suggestions, 39,853 order suggestion items, 3,144 purchase orders, 49,176 purchase order items, and 5,144 materials. - Order suggestion items exported manually earlier than the planner suggested, which locks in less accurate forecasts, occur 1,123 times in the review period and are assessed as a direct source of potential overstock. - Materials created as purchase order items repeatedly without being received, with replacement orders padded by undemanded items to meet minimum quantities or truck capacity, affect 171 materials and are assessed as both shortfall and overstock. - Other patterns, such as batch receipts, partially received orders, single-item purchase orders, and repeated deletion of suggestions, are assessed as shortfall, inefficiency, or rework rather than overstock. - Limitation: the material re-ordering pattern is only visible because objects are linked across purchase orders, which traditional single-case process mining cannot do; the study assumes predicted demand equals actual demand and the proposed measures are not yet operationalized or measured. --- # Leveraging machine learning for automatic topic discovery and forecasting of process mining research: A literature review - Type: Journal article - URL: https://www.gyunam.com/publication/2024-eswa/ - Date: 2024-01-01 - Authors: Gyunam Park, Minsu Cho, Jiyoon Lee - Venue: Expert Systems with Applications - DOI: https://doi.org/10.1016/j.eswa.2023.122435 - PDF: https://www.gyunam.com/publication/2024-eswa/paper.pdf - Tags: Process mining, Text mining, Network analysis, Literature reviews, Emerging trends, Predictions ## Abstract Process mining is a relatively new discipline that focuses on gaining process-centric knowledge from event logs collected by enterprise systems. From an academic standpoint, there has been a constant effort to develop various techniques to automatically discover process models, analyze the compliance of real-life processes to the process models, predict operational frictions, and recommend possible actions to mitigate emerging risks. As far as applications are concerned, process mining techniques have been adopted in various industries, such as healthcare, manufacturing, logistics, and finance. In this work, we analyze the process mining literature in-depth using text mining and machine learning techniques. More in detail, we (1) analyze the main research fields in process mining and their trends, (2) investigate the relationship between the fields, and (3) predict the expansion of the fields in the near future. To that end, we analyze 2,677 process mining articles from 2003 to 2022 using a range of techniques such as topic modeling with the pre-trained language model, exploratory bibliometric analysis, network and community detection, and future prediction. ## Practical problem Researchers, students, and practitioners entering process mining face a literature of thousands of papers spread across many venues and need to know which topics dominate, which are rising, and how they connect. Existing reviews are written by groups of experts who pick topics and judge trends by hand, so the results depend on the reviewers' perspective and offer no systematic way to project where the field is heading. Companies deciding where to invest and tool vendors deciding what to build likewise lack a data-based map of the discipline. A quantitative, reproducible method that treats the literature itself as data can reduce this bias and support forecasting. ## Research questions 1. What are the main research topics of process mining, and how have their trends changed over time? 2. Which process mining topics are mutually cooperative, and how does each research topic relate to the others? 3. How much will process mining expand in the near future, and which research topics will be the focus? ## Contributions - A five-step quantitative literature-review method (data preparation, research-field identification, field investigation, relation identification, future prediction) built on text mining and machine learning. - A corpus of 2,207 process mining articles and proceedings (2002 to 2021) from Web of Science, topic-modeled with BERTopic into 32 research fields grouped into technique, application and support categories. - A keyword-similarity network of research fields (YAKE! keyword extraction, Jaccard similarity) with Louvain community detection to expose relationships between topics. - A Long Short-Term Memory (LSTM) forecasting model fed with a publication representation matrix that combines historical counts and inter-field similarities to predict publications per field. - A comparison of the identified fields with workshop topics at the leading process mining conferences and with the functionalities of tools such as Celonis, UiPath, `ProM` and Disco. ## Key findings - Process discovery (451 articles), conformance checking (136) and frameworks and tools are dominant fields; process monitoring, predictive process monitoring, privacy-preserving process mining and event logs are skyrocketing; process model comparison is deep-cut. - Healthcare is the largest application domain (267 articles); 68% of publications are conference proceedings; the Netherlands and China account for 12.9% and 10.7% of all papers. - Louvain community detection yields four communities (representative, resource-oriented, performance-oriented, manufacturing and Internet of Things) plus six isolated topics; process discovery and frameworks and tools are the central connectors. - The LSTM forecaster reaches a root mean squared error of 5.63 ± 1.02 and mean absolute error of 3.98 ± 0.55 under time-series cross-validation, and predicts growth for most fields in 2022 except variant analysis and frameworks and tools. - Limitations: document preprocessing and field naming still need manual expert steps, and topics the topic model does not surface, notably object-centric process mining, are absent from the trend and forecast analysis. --- # Incorporating Behavioral Recommendations Mined from Event Logs into AI Planning - Type: Conference paper - URL: https://www.gyunam.com/publication/2024-caise-f/ - Date: 2024-01-01 - Authors: Gyunam Park, Majid Rafiei, Hayyan Helal, Gerhard Lakemeyer, Wil M. P. van der Aalst - Venue: 36th International Conference on Advanced Information Systems Engineering Forum, CAiSE Forum2024, Limassol, Cyprus, 03-07 June, 2024, Proceedings - DOI: https://doi.org/10.1007/978-3-031-61000-4_3 - PDF: https://www.gyunam.com/publication/2024-caise-f/paper.pdf - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md), [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract AI planning plays a crucial role in the design and optimization of business processes, providing optimal plans, i.e., sequence of activities, based on manually crafted or formally documented rules. When these plans are executed in business processes, the supporting information systems record a wealth of event data. Analyzing such event data facilitates understanding implicit patterns and recommendations that have the potential to refine planning strategies significantly. In this paper, we introduce a systematic approach to mining these recommendations from event data and integrating them into AI planning, thus creating plans that are informed by both the regulatory hard rules and the flexibility of soft recommendations. ## Practical problem Study advisors, logistics planners and clinicians rely on automated planners to lay out sequences of steps, for example which courses a student should take in which order. Those planners work from hard rules written down in regulations or process models, such as a mandatory prerequisite, and know nothing about what has historically worked well, such as the observation that students who take Algorithms first do better in Advanced Data Structures. That experience sits unused in the event data that information systems record every time a plan is executed. Specifying such preferences by hand is slow and incomplete, so plans stay valid but miss the opportunity to steer people toward behavior that has proven beneficial. ## Research questions 1. How can behavioral recommendations (soft rules) be mined from event logs that record executed plans, and how can their relevance be tested statistically? 2. How can such recommendations be combined with regulatory hard rules inside AI planning so that generated plans are both valid and preferred? ## Contributions - A two-phase approach that mines behavioral recommendations from event data and feeds them into AI planning as soft constraints alongside hard regulatory rules. - A mining procedure that instantiates Declare temporal pattern templates (precedence, response, exclusive choice), checks each case with Linear Temporal Logic (LTL), and keeps only patterns whose satisfied and violated cases differ significantly on an outcome score. - Formalizes recommendation-based planning problems with a weight function that sums the weights of recommended temporal patterns holding in a plan, defining optimal plans as valid plans of maximal weight. - Shows that recommendation-based planning with each action occurring at most once can be solved as an integer linear programming optimization over plan membership and ordering relations. - A running example from curriculum design in which course prerequisites act as hard rules and mined course-ordering patterns act as soft rules. ## Key findings - This is a short vision paper without experiments; it argues that event data from executed plans contains statistically testable behavioral patterns that current planners ignore. - It shows by construction how a Declare pattern such as Precedence(Algorithms, Advanced Data Structures) becomes a weighted preference that raises the rank of plans respecting it while prerequisites remain mandatory. - It establishes that, because any valid plan with non-repeating actions has linear length, the planning problem is in NP and can be encoded as an integer linear program maximizing the recommendation weight. --- # Enhancing healthcare process analysis through object-centric process mining: Transforming OMOP common data models into object-centric event logs - Type: Journal article - URL: https://www.gyunam.com/publication/2024-jbi-omop/ - Date: 2024-01-01 - Authors: Gyunam Park, Yaejin Lee, Minsu Cho - Venue: Journal of Biomedical Informatics - PDF: https://www.gyunam.com/publication/2024-jbi-omop/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Objectives. This study aims to enhance the analysis of healthcare processes by introducing Object-Centric Process Mining (OCPM). By offering a holistic perspective that accounts for the interactions among various objects, OCPM transcends the constraints of conventional patient-centric process mining approaches, ensuring a more detailed and inclusive understanding of healthcare dynamics. Methods. We develop a novel method to transform the Observational Medical Outcomes Partnership Common Data Models (OMOP CDM) into Object-Centric Event Logs (OCELs). First, an OMOP CDM4PM is created from the standard OMOP CDM, focusing on data relevant to generating OCEL and addressing healthcare data’s heterogeneity and standardization challenges. Second, this subset is transformed into OCEL based on specified healthcare criteria, including identifying various object types, clinical activities, and their relationships. The methodology is tested on the MIMIC-IV database to evaluate its effectiveness and utility. Results. The application of our proposed method to the MIMIC-IV dataset effectively produces OCELs, allowing for the implementation of OCPM in the healthcare industry. We rigorously evaluate the comprehensiveness and level of abstraction to validate our approach’s effectiveness. Additionally, we create diverse object-centric process models intricately designed to navigate the complexities inherent in healthcare processes. Conclusion. Our approach introduces a novel perspective by integrating multiple viewpoints simultaneously. To the best of our knowledge, this is the inaugural application of OCPM within the healthcare sector, marking a significant advancement in the field. ## Practical problem Hospitals use process mining to see how patients actually move through care, but these analyses take the patient or the visit as the only case. That hides the parallel workflows of physicians, nurses, pharmacists and medical devices and how they lock together in the same clinical event, such as an X-ray involving a patient, a technologist and a machine at once. Object-centric process mining can show all of these perspectives together, but it needs object-centric event logs, and no one has specified which object types, activity granularities and relationships a clinical log should contain. On top of that, hospital information systems differ in data structure and vocabulary, so results are hard to standardize or reproduce across institutions. ## Research questions 1. How can a standard clinical data model, the Observational Medical Outcomes Partnership Common Data Model (OMOP CDM), be transformed systematically into an object-centric event log? 2. Which object types, levels of clinical activity abstraction and event-to-object relationships should a healthcare object-centric event log contain? 3. Does the transformation preserve data quality and produce object-centric process models that are informative on real hospital data? ## Contributions - OMOP CDM4PM, a formally defined subset of the OMOP CDM that organizes event tables, object tables, event-object relationships and concept relationships for process mining. - A formal transformation from OMOP CDM4PM to an object-centric event log (OCEL) built from identifier, timestamp, event type and object type assignment functions. - A three-level event type assignment (domain, concept class, concept name) and a two-level object type assignment derived from the OMOP concept table and the SNOMED CT hierarchy. - An evaluation protocol that applies the Kahn et al. data quality framework (completeness, conformance) together with abstraction and utility analyses on three MIMIC-IV sub-datasets. - An open-source implementation and manual, `omop2ocel`, released on GitHub (gyunamister/omop2ocel). ## Key findings - On three MIMIC-IV subsets converted to OMOP CDM (100 patients across all visit types, 1,088 inpatients, 913 emergency patients), the transformation keeps every event (364,532; 1,650,306; 1,580,208), object, patient, visit and length-of-stay value unchanged. - Abstracting from low-level to high-level activities cuts the number of events by 85.71%, 95.50% and 95.51% on the three datasets, and the number of event types in the 100-patient set from 2,393 to 25 (middle level) to 7 (high level). - The resulting logs contain eight object types in three groups: patient-related (patients, visits), provider-related (physicians, advanced health professionals, pharmacists) and device-related (dermal drug delivery, injection, diagnostic and supplements). - Object-centric directly-follows graphs over all object types are highly complex, largely because of device types; filtering to providers shows pharmacists handle only drug activities, physicians only condition activities, and advanced health professionals measurements, procedures and observations. - Limitations: no domain experts reviewed the models, the method gives no guidance on choosing an abstraction level, and scalability beyond MIMIC-IV to other institutions remains untested. --- # Analyzing interconnected processes: using object-centric process mining to analyze procurement processes - Type: Journal article - URL: https://www.gyunam.com/publication/2025-ijdsa-p2p/ - Date: 2023-07-16 - Authors: Alessandro Berti, Urszula Jessen, Gyunam Park, Majid Rafiei, Wil M. P. van der Aalst - Venue: International Journal of Data Science and Analytics - DOI: https://doi.org/10.1007/s41060-023-00427-3 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract The purchase-to-pay (P2P) process is one of the core business processes in any organization. It ensures the correct and efficient provisioning of materials and services. An efficient P2P process reduces operational costs by ensuring discounts, avoiding late payments, and choosing the optimal supplier for the goods. Process mining techniques help practitioners optimize the execution of P2P processes by analyzing the execution data and providing useful insights. However, existing techniques may result in misleading insights due to many-to-many relationships between business objects, e.g., between orders and invoices in the P2P process. Recently, object-centric process mining techniques have been proposed to avoid the limitations of traditional process mining techniques. In this paper, we present a case study on a real-life P2P process using object-centric process mining techniques. To that end, we adopt the well-known PM $$^{2}$$ 2 process mining project methodology to the object-centric setting and analyze the performance and compliance. --- # Preserving Complex Object-Centric Graph Structures to Improve Machine Learning Tasks in Process Mining - Type: Journal article - URL: https://www.gyunam.com/publication/2023-eaai/ - Date: 2023-07-15 - Authors: Jan Niklas Adams, Gyunam Park, Wil M. P. van der Aalst - Venue: Engineering Applications of Artificial Intelligence - DOI: https://doi.org/10.1016/j.engappai.2023.106764 - PDF: https://www.gyunam.com/publication/2023-eaai/paper.pdf - Link (URL): https://www.sciencedirect.com/science/article/pii/S095219762300948X - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md), [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) - Tags: Object-centric process mining, Machine learning, Predictive process monitoring ## Abstract Interactions of multiple processes and different objects can be captured using object-centric event data. Object-centric event data represent process executions as event graphs of interacting objects. When applying machine learning techniques to object-centric event data, the event log has to be flattened into sequential process executions used as input for traditional process mining approaches. However, sequentializing the events by flattening removes the graph structure of object-centric event data and, therefore, constitutes an information loss. In this paper, we present a general approach to preserve the graph structures of object-centric event data across machine learning tasks in process mining. We provide two different techniques to preserve these structures depending on the required input format of machine learning techniques: as direct graph encodings or as graph embeddings. We evaluate our contributions by applying three different predictive process monitoring tasks to direct graph encodings, graph embeddings, and flattened event logs. Based on the relative performances, we assess the information contained in the graph structure of object-centric event logs that is, currently, lost in machine learning approaches for traditional process mining. Furthermore, we compare different graph embedding techniques for object-centric event logs to derive recommendations for future research and deployment. To conclude, we assess the improvement that graph embeddings constitute over state-of-the-art approaches that capture object information through specific features. Our contributions improve predictive process monitoring on object-centric event data and quantify the potential performance increases of predictive models. These contributions are especially relevant when looking at real-life applications of predictive process monitoring which are often applied to information systems with relational databases that contain multiple objects and, currently, still enforce flattening. ## Practical problem Companies run their processes on systems such as Enterprise Resource Planning (ERP) software, where one event typically touches several business objects at once: an order, its items, the deliveries and the invoices that belong to it. Analysts who want to predict how such a process will continue (what happens next, when, how long until completion) feed this data into machine-learning models that expect one neat sequence of events per case. Forcing the data into that shape, known as flattening, throws away how objects interact: rework on one object becomes indistinguishable from the same activity on a second object, the number of objects and their individual paths disappear, and an activity can no longer be tied to the object it belonged to. Existing attempts to repair this add a few hand-designed object counts as features, which recover only a small part of what was lost. ## Research questions 1. How much predictive information is lost when the graph structure of object-centric process executions is flattened into event sequences? 2. What is the best way to preserve that graph structure for machine-learning applications in process mining? ## Contributions - A general approach that keeps object-centric process executions as event graphs throughout the machine-learning pipeline, either as direct graph input or as graph embeddings for vector-based models. - Formalizes process executions as event graphs over connected components of the object graph, and flattening as a baseline transformation, with worked examples of three kinds of structural information loss. - An end-to-end predictive pipeline (subgraph extraction, leakage-free train/test split, graph neural network or embedding plus regression/MLP) covering next activity, next timestamp and remaining time. - An evaluation protocol that isolates the value of structure alone (no event features) and a benchmark that measures the added value of embeddings over state-of-the-art object-centric features. - An implementation on top of the `ocpa` Python library with experiments released on GitHub (niklasadams/PreservingOCStructures). ## Key findings - On the BPI Challenge 2017 loan-application log (507,553 events, 67,498 objects, 31,509 process executions), graph-based input beats the flattened baseline for every task and subgraph size k = 2 to 8; the gap usually grows with k. - Averaged over the three prediction tasks, flattening loses structural information worth a 41.7% relative performance improvement with a graph neural network; the best embedding, Family of Graph Spectral Distance (FGSD), reaches 40.8%. - Among seven graph embeddings, FGSD and GL2Vec perform best almost everywhere, NetLSD and Spectral Features worst; MLP slightly beats regression but the embedding choice matters more. - Appended to state-of-the-art object-centric features, FGSD embeddings improve prediction by 2.43% on average across 147 models, and by 4.42% at the largest subgraph size k = 8. - The structure alone carries little value for remaining-time prediction, and the results rest on only two event logs (one synthetic order-management log and one real-life log with two object types). --- # A Generic Approach to Extract Object-Centric Event Data from Databases Supporting SAP ERP - Type: Journal article - URL: https://www.gyunam.com/publication/2023-jiis-generic-sap-extraction/ - Date: 2023-07-13 - Authors: Alessandro Berti, Gyunam Park, Majid Rafiei, Wil MP van der Aalst - Venue: Journal of Intelligent Information Systems - DOI: https://doi.org/10.1007/s10844-023-00799-9 - PDF: https://www.gyunam.com/publication/2023-jiis-generic-sap-extraction/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Process mining provides a collection of techniques to gain insights into business processes by analyzing event logs. Organizations can gain various insights into their business processes by using process mining techniques. Such techniques use event logs extracted from relational databases supporting the business process as input. However, extracting event logs is challenging due to the size of the data, and it remains ad-hoc. Existing commercial tools partly support the extraction of event logs, but they are proprietary and focus on the mainstream processes such as Purchase-To-Pay (P2P) and Order-To-Cash (O2C). Moreover, the extracted event logs suffer from well-known deficiency, convergence, and divergence issues. For example, due to convergence events are unintentionally duplicated causing unreliable or confusing performance diagnostics. In this paper, we propose an approach to extract event logs while avoiding the aforementioned issues. More in detail, we extract object-centric event logs by using an abstraction layer of the database, called Graph of Relationships (GoRs), designing blueprints with domain knowledge, and converting the database and blueprint into object-centric event logs.We fully implemented the proposed approach, which can extract object-centric event logs from SAP ERP systems, and evaluate the utility and scalability of the proposed approach. ## Practical problem Before a company can analyze a process such as purchasing or sales with process mining, someone has to pull an event log out of the database of its Enterprise Resource Planning (ERP) system. In an SAP ERP installation that database holds hundreds of thousands of tables, so analysts struggle to find which tables belong to a process, and writing the extraction queries demands deep domain knowledge. Commercial extractors solve this only for the mainstream Purchase-to-Pay and Order-to-Cash processes, are proprietary, and need customization as soon as a process deviates from the template. Worse, they force every event under a single case identifier, which silently drops events, duplicates them (so performance figures become unreliable), or breaks the real ordering between them. ## Research questions 1. How can the relational structure of an ERP database be abstracted so that analysts can identify processes and select the relevant tables without writing queries? 2. How can table entries be translated into events with activities, timestamps and related objects so that the result is an object-centric event log free of deficiency, convergence and divergence? 3. Is such an extraction feasible in acceptable time on a real SAP ERP instance, and how useful are the resulting logs? ## Contributions - A technology-independent database abstraction (tables, entries, object types, transactions, keys, object relationships) and a derived Graph of Relationships (GoR) that supports process identification and table selection. - The concept of blueprints that map table entries to primitive events, with a basic blueprint needing no domain knowledge and a domain-knowledge blueprint grouping related entries into one event. - A formal conversion from database plus blueprint to an object-centric event log in the OCEL standard, including a projection to a chosen subset of tables. - A concrete realization for SAP ERP: the SQL queries that build the GoR from SAP dictionary tables and a classification of SAP tables into change, transaction and object tables. - The open-source Interactive SAP Explorer tool (Neo4J graph database, Angular front end, Python extractor on GitHub) plus two extracted O2C and P2P logs published in JSON-OCEL. ## Key findings - On an educational SAP ERP instance, the full database abstraction (662,161 tables, 934,546 attributes, 9.6 million attribute arcs, 1.67 million relationship arcs) is built in about 24 seconds of query and postprocessing time. - Extracting entries from Procure-to-Pay tables is fast (1.25 s for 1,844,532 BSEG payment rows), but computing object relationships dominates runtime: 137.79 s of postprocessing for 2,323,792 BSEG relationships. - Starting from the purchasing-document object type EINKBELEG, GoR expansion reaches the expected purchase-requisition, invoice and payment tables, and process identification surfaces twelve processes, e.g. sales documents with 412,228 events across 11 tables. - Object-centric directly-follows graphs discovered from the extracted O2C and P2P logs show the document flows and change activities of both processes without convergence or divergence artifacts. - The basic blueprint yields activity names of lower quality than ad-hoc commercial extractors, the extractor runs only against SAP ECC on Oracle with memory-bound Python extraction, and table-set completeness remains the user's responsibility. --- # Optimizing Resource Allocation Based on Predictive Process Monitoring - Type: Journal article - URL: https://www.gyunam.com/publication/2023-access/ - Date: 2023-06-01 - Authors: Gyunam Park, Minseok Song - Venue: IEEE Access - DOI: https://doi.org/10.1109/ACCESS.2023.3267538 - PDF: https://www.gyunam.com/publication/2023-access/paper.pdf - Link (URL): https://doi.org/10.1109/ACCESS.2023.3267538 - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Recent breakthroughs in predictive business process monitoring equip process analysts with predictions on running process instances, supporting the elicitation of proactive measures to mitigate risks that can be caused by the process instance. However, contrary to active research on providing various predictions and improving the accuracy of prediction models, the practical application of such predictions has been left to the subjective judgment of domain experts. In this work, we explore the exploitation of the insights from predictive information for the actual process improvement in practice. Concretely, we focus on improving resource allocation in business processes where the goal is to allocate appropriate resources to tasks at the proper time. Based on design science methodology, we develop a two-phase method to improve resource allocation by leveraging predictions. Based on the method, we instantiate an algorithm to optimize total-weighted completion time and evaluate its effectiveness and efficiency. From an academic standpoint, our work demonstrates the combination of predictions using machine learning and optimizations based on scheduling. From a practical standpoint, our work provides a general approach to optimize resource allocations for different objectives using predictions. ## Practical problem Hospitals, service centers, and other organizations must constantly decide which staff member handles which waiting case next, and a poor choice means urgent cases wait while a slow worker handles an easy one. Classical scheduling methods can compute good assignments, but only if they know each case's upcoming steps and how long each worker would take, information that is simply not available while a patient is still being treated. Meanwhile, predictive process monitoring has become good at forecasting exactly such missing details, yet in practice its forecasts end up on a dashboard for a manager to interpret rather than driving any decision. Process owners therefore have predictions on one side and allocation algorithms on the other, with no systematic way to connect the two. ## Research questions 1. How can predictions about running process instances be used to optimize resource allocation when the information required by standard scheduling techniques is missing (non-clairvoyant instances)? 2. Does prediction-enriched allocation improve total weighted completion time over the best allocation possible with only known information, and at what computational cost? 3. How accurate do the predictions need to be before the allocation outperforms the no-prediction baseline? ## Contributions - A two-phase method, formalized through scheduling instances, work assignments, schedules and objective functions, that first trains prediction models offline and then predicts missing parameters online before optimizing the allocation. - An uncertainty-aware prediction step based on Bayesian neural networks with Monte Carlo dropout and LSTM layers, which discards predictions whose uncertainty exceeds a threshold. - An instantiation for minimizing total weighted completion time of weight-aware non-clairvoyant instances that casts allocation as a minimum cost maximum flow problem solved by the network simplex method. - An evaluation protocol comparing against a maximum-weight baseline on a simulated emergency department and a real tertiary-hospital emergency log, including a controlled study of prediction accuracy. - A public Python implementation on GitHub (`prediction_based_resource_allocation`). ## Key findings - On the simulated emergency department (11 activities, 25 resources), the method yields lower total weighted completion time than the baseline for every patient set; with 100 patients the baseline is 14% higher (6,348 vs 5,420). - Prediction adds computation, but the simulated 3-hour process is scheduled in under 80 seconds, which the paper judges suitable for real-time use. - Scheduling quality jumps between 40% and 60% emulated prediction accuracy (with 110 patients, TWCT falls from 7670 to 6036); about 60% accuracy usually beats the baseline, with the worst case needing 80% and the best only 50%. - On the real emergency log (421,995 events, 27,478 patients, 754 resources), the method beats the baseline on all 15 daily instance sets; on November 1, 2018 it reaches TWCT 4871 versus 5386, a 9.6% improvement. - Limitations: on the real log almost all run time goes into predictions (710.4 of 721.6 seconds in one day), accuracy and objective are not monotonically related because the future sequence stays unknown, and alternative-resource service times had to be estimated rather than observed. --- # Performance-preserving event log sampling for predictive monitoring - Type: Journal article - URL: https://www.gyunam.com/publication/2023-sampling-jiis/ - Date: 2023-03-06 - Authors: Mohammadreza Fani Sani, Mozhgan Vazifehdoostirani, Gyunam Park, Marco Pegoraro, Sebastiaan J van Zelst, Wil MP van der Aalst - Venue: Journal of Intelligent Information Systems - DOI: https://doi.org/10.1007/s10844-022-00775-9 - PDF: https://www.gyunam.com/publication/2023-sampling-jiis/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Predictive process monitoring is a subfield of process mining that aims to estimate case or event features for running process instances. Such predictions are of significant interest to the process stakeholders. However, most of the state-of-the-art methods for predictive monitoring require the training of complex machine learning models, which is often inefficient. Moreover, most of these methods require a hyper-parameter optimization that requires several repetitions of the training process which is not feasible in many real-life applications. In this paper, we propose an instance selection procedure that allows sampling training process instances for prediction models. We show that our instance selection procedure allows for a significant increase of training speed for next activity and remaining time prediction methods while maintaining reliable levels of prediction accuracy. ## Practical problem Analysts who want to predict the next activity or the remaining time of running cases increasingly rely on machine learning models such as LSTM (Long Short-Term Memory) networks and XGBoost (eXtreme Gradient Boosting). Training these models on a full event log is slow, and getting acceptable accuracy requires repeated training runs for hyperparameter tuning, which is not feasible when models must be refreshed frequently to follow a changing process or when hardware is limited. Generic instance selection techniques from machine learning could shrink the training data, but they assume independent instances, whereas the prefixes extracted from one case are strongly correlated. Practitioners therefore lack a way to cut training time without silently degrading prediction quality. ## Research questions 1. Can prediction models trained on a sampled training log reach similar accuracy to models trained on the whole log while training substantially faster? 2. Which sampling distribution (unique, logarithmic or division) and which log characteristics determine whether accuracy is preserved? 3. Does the sensitivity to sampling differ between next activity prediction and remaining time prediction? ## Contributions - A four-step instance selection procedure for event logs: traverse variants, compute attribute distributions per variant, sort the traces of each variant, and return a sample of the training log only. - Three selection strategies, unique, logarithmic (log base k) and division (ceiling of frequency over k), with a formal notion of variant-preserving sampling. - Relative evaluation metrics for size reduction, accuracy, F1-score, MAE (mean absolute error), RMSE (root mean squared error), feature extraction time and training time against the full-log baseline. - An implementation on both Java and Python platforms, with code and feature generation published at `github.com/gyunamister/pm-prediction`. - An evaluation protocol on three real-life logs (BPIC2012-W, Road Traffic Fine Management, Sepsis) with LSTM and XGBoost, 5-fold cross-validation repeated 5 times on one CPU thread. ## Key findings - Logarithmic sampling with base 10 gives the largest reduction; the Road Traffic Fine Management training log shrinks by more than 1000 times, while division sampling barely reduces Sepsis because most of its variants are unique. - For next activity prediction, division sampling keeps accuracy and F1-score close to the full-log model (relative accuracy near 1) for both LSTM and XGBoost; division by 2 on Sepsis with LSTM even raises accuracy. - Unique and logarithmic sampling distort the variant frequency distribution of Road Traffic Fine Management, which has a few highly frequent variants, and lower accuracy and F1-score; base-10 logarithmic sampling there counts as oversampling. - Training-time speed-up grows with the size reduction and is mostly larger for XGBoost than for LSTM; the choice of distribution matters more than the value of k. - Remaining time prediction is more sensitive to sampling: XGBoost preserves MAE and RMSE only without oversampling, and the LSTM remaining time model is inaccurate even on the full log (high bias), so sampling cannot be judged from it. --- # Monitoring Constraints in Business Processes Using Object-Centric Constraint Graphs - Type: Conference paper - URL: https://www.gyunam.com/publication/2022-pqmi/ - Date: 2023-03-01 - Authors: Gyunam Park, Wil M. P. van der Aalst - Venue: Process Mining Workshops - ICPM 2022 International Workshops, Bozen-Bolzano, Italy, October 23-28, 2022, Revised Selected Papers - DOI: https://doi.org/10.1007/978-3-031-27815-0_35 - PDF: https://www.gyunam.com/publication/2022-pqmi/paper.pdf - Link (URL): https://doi.org/10.1007/978-3-031-27815-0_35 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Constraint monitoring aims to monitor the violation of constraints in business processes, e.g., an invoice should be cleared within 48 h after the corresponding goods receipt, by analyzing event data. Existing techniques for constraint monitoring assume that a single case notion exists in a business process, e.g., a patient in a healthcare process, and each event is associated with the case notion. However, in reality, business processes are object-centric , i.e., multiple case notions (objects) exist, and an event may be associated with multiple objects. For instance, an Order-To-Cash (O2C) process involves order , item , delivery , etc., and they interact when executing an event, e.g., packing multiple items together for a delivery. The existing techniques produce misleading insights when applied to such object-centric business processes. In this work, we propose an approach to monitoring constraints in object-centric business processes. To this end, we introduce Object-Centric Constraint Graphs (OCCGs) to represent constraints that consider the interaction of objects. Next, we evaluate the constraints represented by OCCGs by analyzing Object-Centric Event Logs (OCELs) that store the interaction of different objects in events. We have implemented a web application to support the proposed approach and conducted two case studies using a real-life SAP ERP system. ## Practical problem Compliance officers and process owners define rules such as 'an invoice must be cleared within 48 hours of the goods receipt' or 'clearing an invoice must involve exactly one invoice' and want to be told from event data when the rules are broken. Existing constraint monitoring techniques assume every event belongs to a single case, for example one patient or one order, but in real purchase-to-pay or order-to-cash processes one event touches a purchase order, several materials, a goods receipt and an invoice at once. Squeezing such data into a single case notion produces false violations, for instance reporting that an item was picked after the invoice was sent when the two belong to different object lifecycles. Rules about which and how many objects an activity may involve cannot be expressed at all. ## Research questions 1. How can constraints that account for the interaction of multiple object types be represented formally and graphically? 2. How can violations of such constraints be evaluated directly on object-centric event logs, including constraints on object involvement and cardinality? 3. Is the approach feasible on event data from a real SAP ERP system? ## Contributions - Object-centric behavioral metrics over Object-Centric Event Logs (OCELs): causal, concurrent and choice ordering relations per object type, object involvement metrics (absent, singular, multiple) and underspecified performance metrics. - Object-Centric Constraint Graphs (OCCGs), a directed graph notation with activity, object type and formula nodes and control-flow, object involvement and performance edges with thresholds, plus formal violation semantics. - A monitoring engine that computes the metrics from an OCEL, reusing the OPerA object-centric performance analysis for sojourn, waiting, synchronization, pooling and lagging times, and evaluates OCCG violations. - A web application (`ProPPa`) for importing OCELs in JSON, XML and CSV, designing OCCGs graphically and visualizing monitoring results, with demo video, manual and event logs on GitHub. - Two case studies on a production process and a purchase-to-pay process of an SAP ERP (enterprise resource planning) system. ## Key findings - In the motivating order-to-cash example, flattening on orders makes three control-flow constraints appear violated, whereas the item-aware version of the first constraint is satisfied because each item is picked directly before packing. - In the production process (four object types, monthly logs Jan. to Mar. 2022), 'no reservation for purchase requisition approval' and 'excessive reservations per production order' are violated in all three months, 'skipping approval' only in Feb. and Mar., and 'delayed purchase order release' only in Feb. - In the purchase-to-pay process (five object types, three windows from Aug. 2021 to May 2022), concurrency between verify material and plan goods issue and clearance of multiple invoices are violated only in the first two windows, while excessive materials per order and delayed order creation appear only in the last. - The case studies demonstrate feasibility on a fictitious company's SAP system rather than measuring accuracy against a baseline; OCCGs cover only the constraints introduced, cannot separate direct from eventual causality and lack timed constraints. --- # Interactive Process Identification and Selection from SAP ERP - Type: Conference paper - URL: https://www.gyunam.com/publication/2022-icpm-demo/ - Date: 2023-03-01 - Authors: Julian Weber, Gyunam Park, Majid Rafiei, Wil M. P. van der Aalst - Venue: Proceedings of the ICPM Doctoral Consortium and Demo Track 2022 co-located with 4th International Conference on Process Mining (ICPM 2022), Bolzano, Italy, October, 2022 - PDF: https://www.gyunam.com/publication/2022-icpm-demo/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract SAP ERP is one of the most popular information systems supporting various organizational processes, e.g., O2C and P2P. However, the amount of processes and data contained in SAP ERP is enormous. Thus, the identification of the processes that are contained in a specific SAP instance, and the creation of a list of related tables is a significant challenge. Eventually, one needs to extract an event log for process mining purposes from SAP ERP. This demo paper shows the tool Interactive SAP Explorer that tackles the process identification and selection problem by encoding the relational structure of SAP ERP in a labeled property graph. Our approach allows asking complex process-related queries along with advanced representations of the relational structure. --- # Pattern-based action engine: Generating process management actions using temporal patterns of process-centric problems - Type: Journal article - URL: https://www.gyunam.com/publication/2023-comind-actionengine/ - Date: 2023-01-01 - Authors: Gyunam Park, Daniel Schuster, Wil M. P. van der Aalst - Venue: Comput. Ind. - DOI: https://doi.org/10.1016/J.COMPIND.2023.104020 - PDF: https://www.gyunam.com/publication/2023-comind-actionengine/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract As business environments become more competitive, organizations strive to improve their business processes to reduce costs and increase quality and productivity. As process improvement traditionally embraces manual creative tasks that are time-consuming and labor-intensive, the need for automating it arises. Action-Oriented Process Mining (AOPM) aims to support automated process improvement by leveraging various process mining techniques. To that end, AOPM first monitors the presence of operational constraints, i.e., operational problems, in business processes, e.g., a high waiting time for patients to register. Next, it produces interim management actions designed to address these transient problems by analyzing the monitoring results. For instance, if an excessive waiting time persists for more than a week, the system might recommend dispatching additional resources for the upcoming week. Contrary to the mature process mining support for monitoring operational constraints, the action part is typically missing in today’s process mining tools. In this work, we propose an action engine to support the automatic generation of actions. It analyzes temporal patterns of monitoring results and produces action plans that describe the execution of management actions. We have demonstrated a use case using the data of a Dutch financial institute to evaluate the feasibility of the proposed action engine and conducted experiments to evaluate its effectiveness. ## Practical problem Organizations monitor their processes for operational problems such as long waiting times or a high share of cancelled orders, and process mining tools are good at flagging these problems. Turning the flags into concrete management actions, however, is still done by hand, which is slow and expensive. The few existing action engines treat each problem as an isolated point event and fire actions on simple rules such as 'if the problem is present, trigger the action'. In practice problems stretch over time intervals and follow one another in patterns, and some remedial actions cannot run at the same time, so rule-based engines produce incomplete or conflicting action plans. ## Research questions 1. How can temporal patterns of interval-based operational problems be specified and detected so that management actions are generated automatically and without conflicts? 2. How do the number of actions and the number of conflicts between them affect the computation time and the scheduling performance of the action planner? 3. How do different conflict structures of the same size affect the computation time and scheduling performance of the action planner? ## Contributions - A pattern-based action engine with three phases: domain experts design action graphs, an action generator derives action requirements from monitoring results, and an action planner schedules conflict-free action plans. - Formalizes constraint instances as interval-based data, temporal pattern trees built on Allen's 13 temporal relations, and action graphs that connect patterns to actions with required durations. - An action conflict model (a directed acyclic graph of precedence constraints) and a makespan-minimizing planner cast as parallel machine scheduling with precedence constraints and unlimited machines. - A use case on the BPI Challenge 2017 loan application process in which frequent temporal patterns are discovered from weekly monitoring results with the process mining tool `Cortado`. - An open-source implementation and experiment scripts on GitHub (aopm/pattern-based-action-engine). ## Key findings - In the loan application use case, an action graph with six temporal patterns and seven actions applied weekly over 12 weeks (October to December 2016) yields a conflict-free plan of minimal makespan every week; plan 2 has makespan 4 days, total waiting time 2 and total flow time 14. - Plans differ by week: plan 9 contains all actions with total waiting time 11, plan 10 contains only two actions with no waiting; waiting arises when an action is planned together with a conflicting predecessor (one action waits in 8 of 9 weeks). - Planner runtime grows quadratically with the number of actions (100 to 1,000) and proportionally with the conflict ratio; with 1,000 actions and 90 percent conflicting pairs the runtime is about 250 seconds, and below 200 actions the conflict ratio hardly matters. - Makespan grows linearly with the number of actions and rises with the conflict ratio, with the largest jump between 40 and 50 percent conflicts; for 500 actions, makespan variance across conflict structures of equal size peaks around 50 percent. - Limitation: the use case was not executed inside the organization, so the effect of the actions is unmeasured; the planner ignores action instances already running and plans fixed weekly requirements, which can delay a response by up to a week. --- # Extracting Rules from Event Data for Study Planning - Type: Conference paper - URL: https://www.gyunam.com/publication/2023-icpm-edupm/ - Date: 2023-01-01 - Authors: Majid Rafiei, Duygu Bayrak, Mahsa Pourbafrani, Gyunam Park, Hayyan Helal, Gerhard Lakemeyer, Wil M. P. van der Aalst - Venue: Process Mining Workshops - ICPM 2023 International Workshops, Rome,Italy, October 23-27, 2023, Revised Selected Papers - DOI: https://doi.org/10.1007/978-3-031-56107-8_28 - Research lines: [Neuro-symbolic process learning](https://www.gyunam.com/research-line/neuro-symbolic-process-learning/index.md), [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) --- # Checking Constraints for Object-Centric Process Executions - Type: Conference paper - URL: https://www.gyunam.com/publication/2023-icpm-pqmi/ - Date: 2023-01-01 - Authors: Tian Li, Gyunam Park, Wil M. P. van der Aalst - Venue: Process Mining Workshops - ICPM 2023 International Workshops, Rome,Italy, October 23-27, 2023, Revised Selected Papers - DOI: https://doi.org/10.1007/978-3-031-56107-8_30 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) --- # Analyzing an After-Sales Service Process Using Object-Centric ProcessMining: A Case Study - Type: Conference paper - URL: https://www.gyunam.com/publication/2023-icpm-pqmi-2/ - Date: 2023-01-01 - Authors: Gyunam Park, Sevde Aydin, Cüneyt Ugur, Wil M. P. van der Aalst - Venue: Process Mining Workshops - ICPM 2023 International Workshops, Rome,Italy, October 23-27, 2023, Revised Selected Papers - DOI: https://doi.org/10.1007/978-3-031-56107-8_31 - PDF: https://www.gyunam.com/publication/2023-icpm-pqmi-2/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Process mining, a technique turning event data into business process insights, has traditionally operated on the assumption that each event corresponds to a singular case or object. However, many real-world processes are intertwined with multiple objects, making them object-centric. This paper focuses on the emerging domain of object-centric process mining, highlighting its potential yet underexplored benefits in actual operational scenarios. Through an in-depth case study of Borusan Cat’s after-sales service process, this study emphasizes the capability of object-centric process mining to capture entangled business process details. Utilizing an event log of approximately 65,000 events, our analysis underscores the importance of embracing this paradigm for richer business insights and enhanced operational improvements. ## Practical problem Service companies dispatch field technicians to customer sites through scheduling systems in which one technician handles many schedules a day and a schedule may involve several technicians. Traditional process mining looks at one schedule at a time, so it cannot show how a technician's day unfolds across several jobs, why planned hours are overrun, or how a delay at one site cascades into the next. Object-centric process mining is designed for such intertwined objects, but tools for it have rarely been applied in practice, and managers lack documented real-life evidence of what it reveals and how to run such a project. ## Research questions 1. How can an organization plan and execute an object-centric process mining project on a service process with many-to-many relations between schedules and technicians? 2. What does the interaction between schedules and technicians look like in practice, and where does it deviate from the company's business rules? 3. Where do bottlenecks arise at the intersection of schedules and technicians, and what improvements follow? ## Contributions - An extension of the PM2 process mining project methodology into five stages for object-centric projects: planning, extraction, data preprocessing, mining and analysis, improvement. - An object-centric event log in the OCEL standard extracted from the Oracle database of Borusan Cat's WeKing scheduling application (65,774 events, 6,483 schedules, 173 technicians). - A case study combining object-centric Petri net discovery, compliance rule monitoring and object-centric performance metrics on the after-sales service process. - A technician-day case notion that makes single-viewpoint analysis of technician workload tractable. - Documented improvement actions implemented at the company and proposals for further ones. ## Key findings - After preprocessing (57,601 events, 5,566 schedules, 169 technicians), 4,282 of 5,566 schedules are overwork, meaning actual hours exceed scheduled hours; about 25% of technician-days exceed 12 hours, and days with a HOLD activity are 3% longer. - The discovered object-centric Petri net reveals deviations from the reference model; in 2023 the rule 'one survey per schedule' is violated 2,222 times, 'no hold before arriving on site' 144 times and 'survey only after job completion' 124 times. - Travel from ENROUTE to ONSITE takes 2.02 hours on average (maximum 12.67) against typical 4-hour schedules, and technicians accept schedules on average 1.92 hours after the scheduled start. - Transit can exceed half of a schedule (one 5-hour schedule spent 4 hours travelling), trips back to the office accumulate further travel, and delays propagate to subsequent schedules of the same technician. - The company introduced mobile reminders and proximity-based assignment; schedules with multiple technicians were excluded from the analysis, and the excessive use of the hold status has not yet led to an action. --- # ocpa: A Python library for object-centric process analysis - Type: Journal article - URL: https://www.gyunam.com/publication/2023-softwareimpact/ - Date: 2022-11-01 - Authors: Jan Niklas Adams, Gyunam Park, Wil M. P. van der Aalst - Venue: Softw. Impacts - DOI: https://doi.org/10.1016/j.simpa.2022.100438 - PDF: https://www.gyunam.com/publication/2023-softwareimpact/paper.pdf - Link (URL): https://doi.org/10.1016/j.simpa.2022.100438 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract ocpa is a Python library supporting object-centric process mining. Traditional process mining generate insights for one single process. However, many real-life processes are composed of multiple interacting subprocesses and events may involve multiple objects. Object-centric process mining provides techniques for analyzing multiple interacting processes by generalizing process mining techniques. ocpa contains algorithms for object-centric event log management, process discovery, conformance checking, enhancement, and predictive process monitoring. ocpa is easily integrable with existing solutions as it supports existing object-centric event log standards, provides extensive documentation, and is installable through the Python package installer. ## Practical problem Process mining tools assume that one execution of a process is a sequence of events belonging to a single case. In an ERP (enterprise resource planning) system an order involves several documents, and in a factory many parts are produced separately and assembled later, so events are shared between objects and the behaviour forms a graph rather than a sequence. To use existing tools, analysts flatten the data onto one case notion, which duplicates events, makes events disappear and introduces false dependencies between activities. Techniques that avoid flattening have been published as separate research prototypes, so practitioners and researchers have had no single platform to apply object-centric discovery, conformance checking, performance analysis and prediction together. ## Research questions 1. How can the object-centric process mining techniques proposed in recent years be bundled into one library that works on object-centric event logs without flattening? 2. How can such a library integrate with existing event log standards and the Python ecosystem so that practitioners and researchers can adopt it? ## Contributions - `ocpa`, an open-source Python library (GPL-3.0, GitHub `ocpm/ocpa`, installable with pip, documented on readthedocs, with a Code Ocean capsule) for object-centric process mining. - A data management submodule that imports CSV, JSON-OCEL and XML-OCEL (Object-Centric Event Log) files into an OCEL class, derives process executions, variants and statistics, and exports JSON-OCEL. - Algorithms for discovering object-centric Petri nets and variant graphs, computing precision and fitness, and monitoring control-flow, object involvement and performance constraints. - Model-based object-centric performance analysis (flow, synchronization, pooling and lagging time plus waiting, service and sojourn time) and feature extraction with table, sequence and graph encodings for predictive monitoring. - A visualization submodule with GraphViz rendering of object-centric Petri nets and a dedicated layout algorithm for variant graphs. ## Key findings - The paper is a software description without experiments; it documents version 1.0.0 of the library, built on pandas, numpy, pm4py, networkx and graphviz. - It argues that the library is the first unified platform covering all four process mining tasks on graph-structured event data, implementing techniques from roughly two years of prior research. - It argues the library enables new research questions, such as clustering object-centric data and comparing predictive models, and supports applied use in finance, manufacturing and service sectors as vendors begin to adopt object-centric data. --- # Explainable Predictive Decision Mining for Operational Support - Type: Conference paper - URL: https://www.gyunam.com/publication/2022-icsoc-workshop/ - Date: 2022-11-01 - Authors: Gyunam Park, Aaron Küsters, Mara Tews, Cameron Pitsch, Jonathan Schneider, Wil M. P. van der Aalst - Venue: Service-Oriented Computing - ICSOC 2022 Workshops - ASOCA, AI-PA, FMCIoT, WESOACS 2022, Sevilla, Spain, November 29 - December 2, 2022 Proceedings - DOI: https://doi.org/10.1007/978-3-031-26507-5_6 - PDF: https://www.gyunam.com/publication/2022-icsoc-workshop/paper.pdf - Link (URL): https://doi.org/10.1007/978-3-031-26507-5_6 - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Several decision points exist in business processes (e.g., whether a purchase order needs a manager’s approval or not), and different decisions are made for different process instances based on their characteristics (e.g., a purchase order higher than e500 needs a manager approval). Decision mining in process mining aims to describe/predict the routing of a process instance at a decision point of the process. By predicting the decision, one can take proactive actions to improve the process. For instance, when a bottleneck is developing in one of the possible decisions, one can predict the decision and bypass the bottleneck. However, despite its huge potential for such operational support, existing techniques for decision mining have focused largely on describing decisions but not on predicting them, deploying decision trees to produce logical expressions to explain the decision. In this work, we aim to enhance the predictive capability of decision mining to enable proactive operational support by deploying more advanced machine learning algorithms. Our proposed approach provides explanations of the predicted decisions using SHAP values to support the elicitation of proactive actions. We have implemented a Web application to support the proposed approach and evaluated the approach using the implementation. ## Practical problem Business processes are full of decision points: a purchase order above a threshold needs a manager's approval, a loan above a certain amount needs an advanced assessment. If a process manager could foresee which branch a running case will take, they could act early, for example steering cases away from a branch that is becoming a bottleneck. Existing decision mining tools explain how past decisions were made using decision trees, but trees predict poorly on new cases because they overfit and are unstable, and a wrong prediction leads to the wrong intervention. Managers need both accurate predictions of upcoming decisions and an explanation of which case attributes drive them, so that a sensible counter-action can be derived. ## Research questions 1. Do advanced machine learning algorithms predict decisions at process decision points more accurately than the decision trees used in traditional decision mining? 2. Does the approach provide reliable explanations of the predicted decisions that reflect the actual decision logic? ## Contributions - Formalizes situation tables, decision models as mappings from feature mappings to likelihood distributions over outgoing transitions of a decision point in a labeled Petri net, and SHAP-based explanations of decisions. - A two-phase approach: an offline phase that computes situation tables from historical event logs and estimates decision models, and an online phase that predicts and explains decisions for running instances. - Uses Shapley Additive Explanations (SHAP) to deliver local explanations per running instance and global explanations by aggregating SHAP values across instances. - An open-source web application, `eXdpn` (GitHub `aarkue/eXdpn`), that discovers Petri nets with the inductive miner, builds situation tables from case, event, and performance features, trains four model families, and visualizes explanations. - An evaluation protocol combining 5-fold cross-validated F1 scores on BPI Challenge 2012 and 2019 with a controlled CPN Tools simulation of a Purchase-to-Pay process whose decision logic is known. ## Key findings - On ten decision points from BPI Challenge 2012 (offers) and 2019 (filtered), XGBoost reaches a top-two F1 score at eight points and support vector machines at six, while the decision tree does so only twice (p16 and p11). - Gains are largest on harder decision points: at p4 and p6 of BPI Challenge 2012 the decision tree scores 0.6888 and 0.7545 F1, versus 0.725 and 0.8048 for the neural network. - On easy decision points all algorithms converge (for example 0.9612 for all methods at p16 and up to 1.0000 at p11), so the advantage of advanced models is small there. - In the simulated Purchase-to-Pay process, the four most impactful SHAP features for the customs-hold decision (EU or non-EU origin, total price, number of items) match the decision logic coded in the simulation, indicating reliable explanations. - SHAP explanations are less expressive than the logical rules produced by tree-based decision mining, and the approach leaves feature definition for situation tables and the link from explanations to concrete actions to future work. --- # Detecting Context-Aware Deviations in Process Executions - Type: Conference paper - URL: https://www.gyunam.com/publication/2022-bpm-forum/ - Date: 2022-09-01 - Authors: Gyunam Park, Janik-Vasily Benzin, Wil M. P. van der Aalst - Venue: Business Process Management Forum - BPM 2022 Forum, Münster, Germany, September 11-16, 2022, Proceedings - DOI: https://doi.org/10.1007/978-3-031-16171-1_12 - PDF: https://www.gyunam.com/publication/2022-bpm-forum/paper.pdf - Link (URL): https://doi.org/10.1007/978-3-031-16171-1_12 - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) ## Abstract A deviation detection aims to detect deviating process instances, e.g., patients in the healthcare process and products in the manufacturing process. A business process of an organization is executed in various contextual situations, e.g., a COVID-19 pandemic in the case of hospitals and a lack of semiconductor chip shortage in the case of automobile companies. Thus, context-aware deviation detection is essential to provide relevant insights. However, existing work 1) does not provide a systematic way of incorporating various contexts, 2) is tailored to a specific approach without using an extensive pool of existing deviation detection techniques, and 3) does not distinguish positive and negative contexts that justify and refute deviation, respectively. In this work, we provide a framework to bridge the aforementioned gaps. We have implemented the proposed framework as a web service that can be extended to various contexts and deviation detection methods. We have evaluated the effectiveness of the proposed framework by conducting experiments using 255 different contextual scenarios. ## Practical problem Hospitals, manufacturers, and service companies flag unusual cases in their processes to catch errors, fraud, or quality problems, but a case that looks abnormal in isolation may be perfectly explainable once the circumstances are known: a long wait for admission during a pandemic wave, or a slow order week when half the team is on leave. Conversely, cases that look normal may hide a problem, for instance throughput that is only on target because staff worked the weekend. Existing deviation detection techniques either ignore such context entirely or bake one specific kind of context into one specific algorithm, so an organization cannot combine the detector it trusts with the contextual factors that matter to it. None of them separates context that excuses a deviation from context that casts doubt on an apparently normal case, which is exactly the distinction a process owner needs when deciding what to investigate. ## Research questions 1. How can arbitrary contextual factors be incorporated systematically into existing unsupervised deviation detection methods, rather than building one method per context? 2. How should positive context (which justifies deviations) and negative context (which refutes apparent normality) be modeled and used to revise deviation scores? 3. Does context-aware post-processing improve the classification of deviating and normal traces across different detection methods and contextual scenarios? ## Contributions - Formalizes context-aware deviation detection as a four-class labeling problem combining non-context and context classifications (context-insensitive or context-sensitive, deviating or normal). - A context ontology for deviation detection that merges prior business process context work into intrinsic, relational, and external contexts, and extends it with positive and negative context semantics. - A post-processing framework with four components (deviation detection, context analysis, context linking to events and traces, score revision) that plugs in any score-based detector. - An extensible cloud-based web service, `contect`, on GitHub with a user manual, implementing workload, weekend, waiting time, and capacity utilization contexts and four detectors (Inductive Miner with alignments, Profiles, ADAR, Autoencoder). - An evaluation protocol that injects four deviation types and four contextual scenarios into CPN Tools simulations of an order management process, yielding 225 scenarios per detector. ## Key findings - Across 225 simulated scenarios per detector, context-aware post-processing with grid-searched positive and negative degrees raises accuracy, precision, and recall over the context-unaware baseline for all four detectors. - For the Inductive Miner based detector, accuracy rises from 0.389118 to 0.426846, precision from 0.248691 to 0.293496, and recall from 0.389118 to 0.425686; precision gains are largest for Autoencoder (+0.133433), Profiles (+0.111686) and ADAR (+0.104261). - Average class accuracy decreases slightly for every detector (e.g. 0.326856 to 0.311832 for Inductive), so the gain is not uniform across the four classes. - Summed confusion matrices show the improvement comes from context-sensitive deviating traces: 54,186 of 163,160 are correctly labeled with Inductive and 47,951 of 159,417 with Autoencoder. - Limitation: context-sensitive normal traces are almost never recovered (0 of 21,381 for Inductive, 83 of 21,240 for Autoencoder), results depend on several tunable parameters and on the underlying detector, and external contexts are only measured indirectly. --- # Action-Oriented Process Mining: Bridging the Gap Between Insights and Actions - Type: Journal article - URL: https://www.gyunam.com/publication/2022-prai/ - Date: 2022-07-01 - Authors: Gyunam Park, Wil MP van der Aalst - Venue: Progress in Artificial Intelligence - DOI: https://doi.org/10.1007/s13748-022-00281-7 - PDF: https://www.gyunam.com/publication/2022-prai/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract As business environments become more dynamic and complex, it becomes indispensable for organizations to objectively analyze business processes, monitor the existing and potential operational frictions, and take proactive actions to mitigate risks and improve performances. Process mining provides techniques to extract insightful knowledge of business processes from event data collected during the execution of the processes. Besides, various approaches have been suggested to support the real-time (predictive) monitoring of the process-related problems. However, the link between the insights from the continuous monitoring and the concrete management actions for the actual process improvement is missing. Action-oriented process mining aims at connecting the knowledge extracted from event data to actions. In this work, we propose a general framework for action-oriented process mining covering the continuous monitoring of operational processes and the automated execution of management actions. Based on the framework, we suggest a cube-based action engine where actions are generated by analyzing monitoring results in a multi-dimensional way. The framework is implemented as a ProM plug-in and evaluated by conducting experiments on both artificial and real-life information systems. ## Practical problem Companies run standardized processes such as order-to-cash, yet real executions produce thousands of variants and recurring frictions: late deliveries, slow quotations, orders changed after confirmation. Process mining makes these problems visible, and monitoring techniques detect or predict them in near real time, but the step from a detected problem to a management action is left to ad-hoc human judgment, and commercial action engines generate actions without a systematic method. Process owners therefore watch dashboards full of violations while the actual interventions, such as reprioritizing an order, alerting a manager or adding a resource, stay manual, slow and inconsistent. What is missing is a structured way to turn a continuous stream of monitoring results into actions that the underlying information system executes automatically. ## Research questions 1. How can the continuous monitoring of operational processes and the automated execution of management actions be connected in one general framework? 2. How can a stream of constraint violations be structured and analysed at suitable abstraction levels so that the needed management actions are generated? 3. Do the monitor and the action engine detect violations and generate transactions correctly, and does executing those transactions in a real information system improve the process? ## Contributions - A general framework for action-oriented process mining with a constraint monitor that turns an event stream into a constraint instance stream and an action engine that turns it into an action instance stream. - Formal definitions of constraints, time moments, constraint instances, transactions, actions and action instances over object-centric event data in which multiple object types interact. - A cube-based action engine: constraint cubes with constraint, context and time dimensions, OLAP (online analytical processing) operations such as slice, dice, roll-up and drill-down, assessment functions and transaction mappings. - An open-source `ProM` plug-in, `ActionOrientedProcessMining`, that reads JSON-OCEL (Object-Centric Event Log) streams and defines constraints and actions with a Constraint Formula Language and an Action Formula Language; manual and data on GitHub. - An evaluation protocol that drives a real SAP ERP (enterprise resource planning) ECC 6.0 system with CPN Tools-simulated automation logs and executes generated actions through a dedicated gateway. ## Key findings - On the artificial order-handling system (16 resources, 72-hour delivery agreement), the monitor detects every late order at each 24-hour check and the engine issues the specified priority raises and manager notifications for the 40 inspected orders. - Over 30 simulated days the number of orders violating the 72-hour agreement is lower at every point in time when the generated actions are applied than when they are not. - In the SAP order-to-cash process (151 inquiries), alerting sales managers and adding a temporary resource when late responses exceed 10 percent lowers mean response time from 30.10 hours (2 Dec.) to 23.35 hours (7 Dec.) and its standard deviation from 17.88 (4 Dec.) to 1.25. - Notifying customers and inserting a sales-manager approval step cut orders with multiple post-confirmation changes from 18 percent to 4 of 53 changed orders (7.5 percent). - The framework is not evaluated with real human participants, has no feedback mechanism to catch conflicting or counter-productive actions, and reduces constraint outcomes to a binary OK/NOK. --- # Analyzing Process-Aware Information System Updates Using Digital Twins of Organizations - Type: Conference paper - URL: https://www.gyunam.com/publication/2022-rcis/ - Date: 2022-05-01 - Authors: Gyunam Park, Marco Comuzzi, Wil M. P. van der Aalst - Venue: Research Challenges in Information Science - 16th International Conference, RCIS 2022, Barcelona, Spain, May 17-20, 2022, Proceedings - DOI: https://doi.org/10.1007/978-3-031-05760-1_10 - PDF: https://www.gyunam.com/publication/2022-rcis/paper.pdf - Link (URL): https://doi.org/10.1007/978-3-031-05760-1_10 - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Digital transformation often entails small-scale changes to information systems supporting the execution of business processes. These changes may increase the operational frictions in process execution, which decreases the process performance. The contributions in the literature providing support to the tracking and impact analysis of small-scale changes are limited in scope and functionality. In this paper, we use the recently developed Digital Twins of Organizations (DTOs) to assess the impact of (process-aware) information systems updates. More in detail, we model the updates using the configuration of DTOs and quantitatively assess different types of impacts of information system updates (structural, operational, and performance-related). We implemented a prototype of the proposed approach. Moreover, we discuss a case study involving a standard ERP procure-to-pay business process. ## Practical problem Companies keep making small changes to the ERP and CRM systems that run their business processes, such as tightening a business rule or adding a field to an order, and digital transformation makes these updates more frequent. Because such systems are woven deeply into daily operations, it is hard to tell in advance what a change touches: which business objects and functions, which running orders or invoices, and how performance measures like waiting time and throughput will shift. Existing impact analysis approaches for ERP changes are model-driven, narrow in scope and lack working implementations, so analysts assess updates largely by hand. ## Research questions 1. How can updates of a process-aware information system be modeled as configuration changes of a digital twin of an organization? 2. How can the structural, operational and performance impacts of such updates be quantified from the digital twin? 3. Is the approach feasible as a software artifact applied to an ERP-style procure-to-pay process? ## Contributions - A meta-model of process-aware information system (PAIS) entities, updates to business functions and business objects, and three impact types: structural, operational and performance. - Formalizes PAIS updates as actions and action instances that change the configuration (valve and operation assignments) of a Digital Twin Interface Model (DT-IM) built on object-centric Petri nets, with a notion of effective change. - Formal definitions of impacted object types and transitions, of the running objects affected via markings and a transition relation function, and of performance impact as the difference in diagnostics before and after an update. - A Web-based prototype, `impacta` (GitHub gyunamister/impacta), that discovers an object-centric Petri net from OCEL data, synchronizes configuration and state via token-based replay, executes actions and reports impacts. - A case study on a simulated procure-to-pay PAIS modeled after the standard SAP ERP process. ## Key findings - In a simulated procure-to-pay system with 24 resources and five object types, raising the minimum order quantity valve to 10 affects one business function (create purchase order) and 51 running objects; average total service time of purchase orders falls by 4 minutes, the sojourn time of create purchase order by 9 minutes, and 13 fewer purchase orders are created. - Changing the verify materials operation to also write a quality attribute affects one object type (material) and one function, touching 254 material objects and 153 objects of verify material; average total service time of purchase orders rises by 1.4 hours and the sojourn time of verify material by 1.6 hours. - The prototype demonstrates the full loop of designing a DT-IM, scheduling action instances on a timeline, and computing structural and operational impacts at action start and performance impacts at action end. - The evaluation is limited to feasibility on an artificial system; no study with business analysts was conducted. - Limitations: only objects already in the process are counted, objects that may bypass an impacted transition are included, and performance impact is measured over all objects rather than isolated to those affected. --- # OPerA: Object-Centric Performance Analysis - Type: Conference paper - URL: https://www.gyunam.com/publication/2022-er/ - Date: 2022-01-01 - Authors: Gyunam Park, Jan Niklas Adams, Wil M. P. van der Aalst - Venue: Conceptual Modeling - 41st International Conference, ER 2022, Hyderabad, India, October 17-20, 2022, Proceedings - DOI: https://doi.org/10.1007/978-3-031-17995-2_20 - PDF: https://www.gyunam.com/publication/2022-er/paper.pdf - Link (URL): https://doi.org/10.1007/978-3-031-17995-2_20 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Performance analysis in process mining aims to provide insights on the performance of a business process by using a process model as a formal representation of the process. Existing techniques for performance analysis assume that a single case notion exists in a business process (e.g., a patient in healthcare process). However, in reality, different objects might interact (e.g., order, delivery, and invoice in an O2C process). In such a setting, traditional techniques may yield misleading or even incorrect insights on performance metrics such as waiting time. More importantly, by considering the interaction between objects, we can define object-centric performance metrics such as synchronization time, pooling time, and lagging time. In this work, we propose a novel approach to performance analysis considering multiple case notions by using object-centric Petri nets as formal representations of business processes. The proposed approach correctly computes existing performance metrics, while supporting the derivation of newly-introduced object-centric performance metrics. We have implemented the approach as a web application and conducted a case study based on a real-life loan application process. ## Practical problem Process analysts measure where time is lost by replaying event data on a process model and reading off waiting, service and sojourn times. These techniques assume every event belongs to exactly one case, but in processes run on systems such as SAP an order, several deliveries and invoices interact and share events. Forcing such data into one case notion duplicates or drops events, so the reported frequencies and durations are simply wrong, and the time spent waiting for several objects to come together cannot be measured at all. Analysts may therefore chase the wrong bottleneck. ## Research questions 1. How can existing performance measures such as waiting, service and sojourn time be computed correctly when process instances involve several interacting object types? 2. Which new performance measures capture the interaction between objects, and how can they be computed on object-centric Petri nets? ## Contributions - An approach that discovers an object-centric Petri net (OCPN) from an object-centric event log (OCEL), replays the log to obtain event occurrences and token visits, and computes performance per event occurrence. - Formalizes replay (event occurrences, token visits) and the relation between an event occurrence and the latest token visit of each object it involves. - Seven measurement functions: flow, sojourn, waiting and service time plus the object-centric synchronization, pooling and lagging time. - A web application implementation (`OPerA`, on GitHub at gyunamister/OPerA) with demo video and manuals. - A case study on the BPI Challenge 2017 loan application log with application and offer object types. ## Key findings - On the filtered loan log (20,478 events, 1,682 applications, 3,573 offers), the flattened analysis with Inductive Visual Miner in `ProM` reports Send repeated by 1,799 applications with average sojourn about 2 days 23 hours, whereas the object-centric view shows it once per offer (92 offers skip it) with about 15 minutes. - Flattening counts Cancel application 1,891 times with average sojourn about 12 days 22 hours; the object-centric analysis yields 1,682 occurrences, once per application, with about 31 days 22 hours. - The new measures show Cancel application has an average synchronization time of about 4 days 11 hours; lagging time is 3 days 15 hours for applications and 19 hours for offers, and pooling time of offers nearly equals synchronization time, so the second offer arrives about 4 days 11 hours after the first. - Limitation: results depend on the quality of the discovered model, and non-conforming behavior can yield misleading insights; model-independent analysis is left to future work. --- # A Framework for Extracting and Encoding Features from Object-Centric Event Data - Type: Conference paper - URL: https://www.gyunam.com/publication/2022-icsoc/ - Date: 2022-01-01 - Authors: Jan Niklas Adams, Gyunam Park, Sergej Levich, Daniel Schuster, Wil M. P. van der Aalst - Venue: Service-Oriented Computing - 20th International Conference, ICSOC 2022, Seville, Spain, November 29 - December 2, 2022, Proceedings - DOI: https://doi.org/10.1007/978-3-031-20984-0_3 - PDF: https://www.gyunam.com/publication/2022-icsoc/paper.pdf - Link (URL): https://doi.org/10.1007/978-3-031-20984-0_3 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Traditional process mining techniques take event data as input where each event is associated with exactly one object. An object represents the instantiation of a process. Object-centric event data contain events associated with multiple objects expressing the interaction of multiple processes. As traditional process mining techniques assume events associated with exactly one object, these techniques cannot be applied to object-centric event data. To use traditional process mining techniques, object-centric event data are flattened by removing all object references but one. The flattening process is lossy, leading to inaccurate features extracted from flattened data. Furthermore, the graph-like structure of object-centric event data is lost when flattening. In this paper, we introduce a general framework for extracting and encoding features from object-centric event data. We calculate features natively on the object-centric event data, leading to accurate measures. Furthermore, we provide three encodings for these features: tabular, sequential, and graph-based. While tabular and sequential encodings have been heavily used in process mining, the graph-based encoding is a new technique preserving the structure of the object-centric event data. We provide six use cases: a visualization and a prediction use case for each of the three encodings. We use explainable AI in the prediction use cases to show the utility of both the object-centric features and the structure of the sequential and graph-based encoding for a predictive model. ## Practical problem Companies that want to predict how long an order will still take, cluster similar executions or chart process performance over time first have to turn their event data into features. The established feature frameworks assume every event belongs to exactly one case, but data from ERP (enterprise resource planning) systems links one event to an order, several items and an invoice at once. To use the existing frameworks, analysts flatten the data onto one object type, which drops events, duplicates them and invents orderings between activities that were actually independent, so features such as 'preceding activity' or 'elapsed time' come out wrong. Flattening also discards the graph shape of the data, which no tabular or sequential encoding can represent, so predictive models never see how objects interact. ## Research questions 1. How can the features of the established process enhancement framework be computed natively on object-centric event data so that they are accurate rather than distorted by flattening? 2. How can object-centric features be encoded as tables, sequences and graphs so that different visualization and prediction techniques can consume them? 3. Do object-centric features and the structure preserved by sequential and graph encodings carry information that predictive models actually use? ## Contributions - A general framework that translates the control-flow, data-flow, resource and performance features of de Leoni et al. to object-centric event logs and adds an object perspective and object-centric performance features. - Formal definitions of the object graph, process executions and two extraction techniques, connected components and leading object type, on Object-Centric Event Logs (OCELs). - Three formally defined feature encodings, tabular, sequential and a new graph-based encoding whose nodes are events and whose edges follow each object's lifecycle. - Six use cases on the BPI Challenge 2017 loan application log, one visualization and one remaining-time prediction per encoding, with SHAP values to explain feature and structure importance. - An open-source implementation in the `ocpa` Python library and a GitHub repository to reproduce the experiments. ## Key findings - A qualitative analysis of 25 features classifies most control-flow, data-flow and performance features as misleading when computed on flattened data, while synchronization, pooling, lagging time and object-count features exist only on object-centric data. - Predicting normalized remaining time on the loan log with the same four features, test MAE (mean absolute error) is 0.5087 for linear regression on the tabular encoding, 0.4568 for an LSTM (Long Short-Term Memory) network on sequences and 0.4497 for a graph neural network, against 0.7598 for predicting the training mean. - SHAP values of the regression model show that a Call activity among preceding activities and a higher number of previous offer objects both lower the predicted remaining time, so the new object feature is informative. - For the LSTM, features at all four positions of the input sequence carry high SHAP importance, and for the graph neural network edge importances differ substantially across edges, so both encodings' structure is used by the models. - The evaluation is illustrative rather than a benchmark: one log, a single 0.7/0.3 split, inputs truncated to four events, and the sequential encoding still imposes precedence relations that the log does not contain. --- # Event Log Sampling for Predictive Monitoring - Type: Conference paper - URL: https://www.gyunam.com/publication/2021-ml4pm/ - Date: 2021-11-01 - Authors: Mohammadreza Fani Sani, Mozhgan Vazifehdoostirani, Gyunam Park, Marco Pegoraro, Sebastiaan J. van Zelst, and Wil M.P. van der Aalst - Venue: International Workshop on Leveraging Machine Learning in Process Mining (ML4PM) - DOI: https://doi.org/10.1007/978-3-030-98581-3_12 - PDF: https://www.gyunam.com/publication/2021-ml4pm/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) ## Abstract Predictive process monitoring is a subfield of process mining that aims to estimate case or event features for running process instances. Such predictions are of significant interest to the process stakeholders. However, state-of-the-art methods for predictive monitoring require the training of complex machine learning models, which is often inefficient. This paper proposes an instance selection procedure that allows sampling training process instances for prediction models. We show that our sampling method allows for a significant increase of training speed for next activity prediction methods while maintaining reliable levels of prediction accuracy. ## Practical problem Analysts who forecast the next step of running cases train deep neural networks on event logs, and training can take hours to weeks. They must repeat it many times to tune parameters and again whenever the process changes, so the long training times limit how often models can be refreshed and on what hardware. Generic data-reduction techniques from machine learning pick a subset of training examples, but they assume examples are independent, whereas events in a log belong to cases and are strongly correlated. A way to shrink an event log for training that respects its case structure is missing. ## Research questions 1. Can training predictive monitoring models on a sampled event log speed up training while keeping accuracy close to that of the full log? 2. Is there a relationship between event log characteristics and the suitable sampling strategy and size? ## Contributions - An event log sampling procedure for predictive monitoring that traverses variants, computes attribute distributions per variant, sorts cases within each variant and returns a subset of cases. - Three variant-aware selection rules: unique (one case per variant), logarithmic (log base k of the variant frequency) and division (frequency divided by k), the last guaranteed to preserve all variants. - Formalizes variant-preserving sampling and the evaluation measures reduction rate, relative accuracy and training and feature-extraction speed-up. - An implementation of the sampling methods in `ProM`, with feature extraction and prediction code on GitHub. - An evaluation protocol on two real logs (Road Traffic Fine Management and BPI Challenge 2012 W) with LSTM and XGBoost next-activity predictors, 5-fold cross-validation repeated five times. ## Key findings - Division sampling preserves accuracy: on the Road log, d10 shrinks the log 9.75 times and speeds up LSTM training 9.0 times and XGBoost training 84.1 times while relative accuracy stays at 0.990 and 1.000; d3 even raises LSTM accuracy by 0.4 percent. - Aggressive sampling trades accuracy for speed: on Road, log10 gives a 572.29-fold reduction and 355.0 times faster XGBoost training but relative accuracy drops to 0.772 (XGBoost) and 0.767 (LSTM); unique sampling falls to 0.582 and 0.631. - Feature-extraction time shrinks more than proportionally to log size: the 572-fold log10 reduction on Road yields a 74,912-fold feature-extraction speed-up. - The choice of distribution matters more than k and depends on the log: BPIC-2012-W, with many unique variants, barely shrinks under division (d2 reduction 1.22) but tolerates unique sampling (LSTM relative accuracy 0.996). - Limitation: very high reduction makes accuracy unacceptable, so sample size should be adjusted gradually in hyperparameter optimization, and the evidence rests on only two logs. --- # Realizing A Digital Twin of An Organization Using Action-oriented Process Mining - Type: Conference paper - URL: https://www.gyunam.com/publication/2021-icpm/ - Date: 2021-11-01 - Authors: Gyunam Park, Wil M.P. van der Aalst - Venue: 2021 International Conference on Process Mining (ICPM) - DOI: https://doi.org/10.1109/ICPM53251.2021.9576846 - PDF: https://www.gyunam.com/publication/2021-icpm/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) - Tags: Conference Proceedings, Action-Oriented Process Mining, ICPM2021 ## Abstract A Digital Twin of an Organization (DTO) is a mirrored representation of an organization, aiming to improve the business process of the organization by providing a transparent view over the process and automating management actions to deal with existing and potential risks. Unlike wide applications of digital twins to product design and predictive maintenance, no concrete realizations of DTOs for business process improvement have been studied. In this work, we aim to realize DTOs using action-oriented process mining, a collection of techniques to evaluate violations of constraints and produce the required actions. To this end, we suggest a digital twin interface model as a transparent representation of an organization describing the current state of business processes and possible configurations in underlying information systems. By interacting with the representation, process analysts can elicit constraints and actions that will be continuously monitored and triggered by an action engine to improve business processes. We have implemented a web service to support it and evaluated the feasibility of the proposed approach by conducting a case study using an artificial information system supporting an order handling process. ## Practical problem Organizations must keep improving processes such as order handling, but with tens of thousands of process variants running on systems like SAP ERP, analysts struggle to see the current state of a process, decide which corrective action to take, and apply it in time. Digital twins promise a live mirror of reality, yet concrete realizations exist only for physical assets such as vehicles and production lines, not for improving business processes across an organization. Monitoring techniques can detect delays and rule violations, but they stop at diagnostics and leave the configuration of the information system to be changed by hand. Process owners need a representation that shows both the state of the process and the levers available in the system, and a mechanism that pulls those levers automatically when a problem recurs. ## Research questions 1. How can a digital twin of an organization be represented so that it exposes both the current state of a business process and the configurable options of the supporting information system? 2. How can process analysts turn recurring problems into action patterns whose constraints are monitored and whose actions are triggered automatically? 3. Is it feasible to realize such a digital twin with action-oriented process mining on a running information system? ## Contributions - A digital twin interface model (DT-IM) formalized as an object-centric Petri net extended with valves (configurable system parameters) and guards encoding routing and resource allocation rules. - Formal definitions of the operational view (marking, object values, diagnostics) and the control view (configurations and actions), together with an operation engine and an action engine. - Taxonomies of constraint violations (compliance- versus performance-oriented) and management actions (inflow, routing, resource allocation, capacity) to support eliciting action patterns. - A containerized web service (`dtween`, on GitHub) that builds DT-IMs from Object-Centric Event Log (OCEL) data, updates states by token-based replay, visualizes them, and evaluates action patterns. - A proof-of-concept case study with a publicly available artificial information system that supports an order handling process with 14 resources. ## Key findings - A DT-IM of the order handling system is discovered from an OCEL; the operational view reports a 16-hour average sojourn time for send notification and a 0.3 routing probability for check availability with approval over two days. - Two action patterns (skip more notifications when the two-day average sojourn time exceeds 16 hours, skip fewer when it falls below 10 hours) are evaluated every 12 hours over 24 days of streaming events in May 2021. - Action A1 fires on 6, 18 and 25 May 2021 and sets the sn-price valve to 220, after which the sojourn time falls until the relaxed constraint triggers A2, which resets it to 180. - With actions, the average sojourn time at 12:00 on 10 May 2021 is 9 hours versus 20 hours without actions. - The evaluation uses an artificial system only; applicability to real-life information systems, real-time state updates, and inter-effects between actions remain unevaluated. --- # An Event Data Extraction Approach from SAP ERP for Process Mining - Type: Conference paper - URL: https://www.gyunam.com/publication/2021-pqmi/ - Date: 2021-11-01 - Authors: Alessandro Berti, Gyunam Park, Majid Rafiei, and Wil van der Aalst - Venue: 6th International Workshop on Process Querying, Manipulation, and Intelligence (PQMI) - DOI: https://doi.org/10.1007/978-3-030-98581-3_19 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract The extraction, transformation, and loading of event logs from information systems is the first and the most expensive step in process mining. In particular, extracting event logs from popular ERP systems such as SAP poses major challenges, given the size and the structure of the data. Open-source support for ETL is scarce, while commercial process mining vendors maintain connectors to ERP systems supporting ETL of a limited number of business processes in an ad-hoc manner. In this paper, we propose an approach to facilitate event data extraction from SAP ERP systems. In the proposed approach, we store event data in the format of object-centric event logs that efficiently describe executions of business processes supported by ERP systems. To evaluate the feasibility of the proposed approach, we have developed a tool implementing it and conducted case studies with a real-life SAP ERP system --- # Discovery of Resource-Oriented Transition Systems for Yield Enhancement in Semiconductor Manufacturing - Type: Journal article - URL: https://www.gyunam.com/publication/2020-ieee-tsm/ - Date: 2021-09-24 - Authors: Minsu Cho, Gyunam Park, Minseok Song, Jinyoun Lee, Byeongeon Lee, Euiseok Kum - Venue: IEEE Transactions on Semiconductor Manufacturing - DOI: https://doi.org/10.1109/TSM.2020.3045686 - PDF: https://www.gyunam.com/publication/2020-ieee-tsm/paper.pdf - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) - Tags: Journal Article, Manufacturing, IEEE-TSM ## Abstract In semiconductor manufacturing, data-driven methodologies have enabled the resolution of various issues, particularly yield management and enhancement. Yield, one of the crucial key performance indicators in semiconductor manufacturing, is mostly affected by production resources, i.e., equipment involved in the process. There is a lot of research on finding the correlation between yield and the status of resources. However, in general, multiple resources are engaged in production processes, which may cause multicollinearity among resources. Therefore, it is important to discover resource paths that are positively or negatively associated with yield. This article proposes a systematic methodology for discovering a resource-oriented transition system model in a semiconductor manufacturing process to identify resource paths resulting in high and low yield. The proposed method is based on the model-based analysis (i.e., finite state machine mining) in process mining and statistical analyses. We conducted an empirical study with real-life data from one of the leading semiconductor manufacturing companies to validate the proposed approach. ## Practical problem In a semiconductor fab a wafer passes through hundreds of process steps, and at each step one of many interchangeable machines or chambers can do the work. Yield engineers want to know which equipment is dragging down yield, but the usual analysis correlates yield with one machine at a time. Because thousands of machines are combined in countless sequences, a chamber that looks excellent in isolation may sit on a path that produces poor wafers, and the single-resource view is confounded by this multicollinearity. Engineers therefore need a way to see which sequences of resources, not individual machines, are associated with high or low yield, in a form compact enough to discuss and act upon. ## Research questions 1. How can resource paths, that is sequences of machines or chambers across process steps, that lead to high or low wafer yield be discovered from manufacturing execution data? 2. Can a model discovered from such paths be simplified to the yield-relevant resources and used to predict the yield class of new wafers? ## Contributions - Formalizes a yield-enhanced event log in which each case (wafer) carries a mandatory yield attribute, and a resource-oriented finite state machine (FSM) whose states and events are the resources of process steps. - A four-phase methodology: data preparation, FSM construction with best- and worst-case paths overlaid, statistical model simplification, and evaluation and interpretation. - A best- and worst-case identification scheme that selects a normality test (Kolmogorov-Smirnov) and then uses either mean and standard deviation or median and median absolute deviation bands with a user-defined weight. - A statistical simplification pipeline (ANOVA with linear contrasts, or Kruskal-Wallis with Mann-Whitney U tests) that labels resources high-class, low-class, or insignificant and merges insignificant resources per step with three merging variants. - An empirical study on real manufacturing execution system data from one of the largest Korean semiconductor manufacturers, with a confusion-matrix evaluation protocol for best, moderate, and worst cases. ## Key findings - On a process of more than 500 steps, 74 expert-selected steps with roughly 13,000 wafers, 950,000 events, and over 1,500 resources are reduced to a model of 74 columns showing only yield-significant resources. - With weight 0.5, 30% of wafers are classified as best and 30% as worst cases; in the first step alone 27% of outgoing arcs are associated with high yield and 36% with low yield. - On the 10% held-out test set the discovered model classifies 83.3% of wafers correctly into best, moderate, or worst; 318 of 393 actual best cases (80.9%) are recovered and worst cases reach 85.7% accuracy. - Almost no wafers are completely reversed (best predicted as worst or vice versa), and domain experts state willingness to use the resource paths for process operation. - Experts flag false negatives, wafers predicted moderate that are actually worst, as the critical residual error needing post-hoc analysis; the method is offline and does not yet use real-time or sensor data. --- # Towards Reliable Business Process Simulation: A Framework to Integrate ERP Systems - Type: Conference paper - URL: https://www.gyunam.com/publication/2021-bpmds/ - Date: 2021-07-17 - Authors: Gyunam Park, Wil M.P. van der Aalst - Venue: International Conference on Business Process Modeling, Development and Support - DOI: https://doi.org/10.1007/978-3-030-79186-5_8 - PDF: https://www.gyunam.com/publication/2021-bpmds/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) - Tags: Conference Proceedings, Simulation, BPMDS2021 ## Abstract A digital twin of an organization (DTO) is a digital replication of an organization used to analyze weaknesses in business processes and support operational decision-making by simulating different scenarios. As a key enabling technology of DTO, business process simulation provides techniques to design and implement simulation models that replicate real-life business processes. Existing approaches have been focusing on providing highly flexible design tools and data-driven evidence to improve the accuracy of simulation models. Provided with such tools and evidence, business analysts are required to reflect comprehensive aspects of reality with subjective judgments, including the design of ERP systems and the organizational interaction with the system. However, given the complexity of ERP systems, it is infeasible and error-prone to manually incorporate the business logic and data restrictions of the system into simulation models, impeding the faithfulness and reliability of the following analysis. In this work, we propose a framework to integrate ERP systems in business process simulation to overcome this limitation and ensure the reliability of the simulation results. The framework is implemented in ProM using the SAP ERP system and CPN Tools. ## Practical problem Business analysts build simulation models to test what-if scenarios before changing a process, for instance adding staff or raising machine capacity, because experimenting on the real organization is costly or risky. The value of such a study depends on how faithfully the model reflects the enterprise system that runs the process, yet systems such as SAP enforce thousands of business rules and data restrictions (SAP alone can raise more than 2,821 distinct errors) that modelers would have to reproduce by hand. In practice models omit behavior the system requires or simulate behavior the system would reject, so the simulated event data and the conclusions drawn from them cannot be trusted. Analysts need a way to run simulations that respect the real system's logic without re-modeling that logic themselves. ## Research questions 1. How can an ERP system be integrated into business process simulation so that its business logic and data restrictions are respected without being modeled in the simulation model? 2. Does executing simulated behavior directly in the ERP system yield event data free of omission and commission problems that can be analyzed like real-life event data? ## Contributions - A framework with three components: an Enterprise Resource Planning (ERP) system abstracted as an object model updated by transactions, a simulation engine producing commands, and a transformation engine translating commands into executions. - Formal definitions of object models, transactions, executions, commands, transaction mappings and parameter mappings that underpin the integration. - An implementation using SAP ERP ECC 6.0 as the system, CPN Tools (Colored Petri Nets) as the simulation engine, and a `ProM` plug-in (ERPSimulator package) that converts XML commands into SAP Remote Function Calls. - A proof of concept that simulates three order-to-cash scenarios (low conversion rate, manual price changes, order cancellations), with CPN files, commands and a manual released on GitHub. ## Key findings - Across the three scenarios, all 288, 4,249 and 4,540 generated commands are translated into Remote Function Calls and executed in SAP without errors, creating 286, 4,093 and 4,384 objects of up to 8 object types. - In the low-conversion scenario, 34 of 73 quotations become orders (46.6%) and every quotation answered later than 10 days is rejected, a pattern recovered from the simulated data with the Inductive Visual Miner in `ProM`. - In the price-change scenario, 113 of 402 orders need manual price changes (79% of the 138 orders containing product P-109); in the cancellation scenario 97 orders are canceled, most often for a delivery date set too late. - The framework assumes a one-to-one mapping between simulated activities and ERP transactions, resolved manually in the implementation; many-to-many relationships and a feedback loop from simulated data to the simulation model remain future work. --- # OCEL: A Standard for Object-Centric Event Logs - Type: Conference paper - URL: https://www.gyunam.com/publication/2021-simpda/ - Date: 2021-07-17 - Authors: Anahita Farhang Ghahfarokhi, Gyunam Park, Alessandro Berti, Wil M.P. van der Aalst - Venue: European Conference on Advances in Databases and Information Systems - DOI: https://doi.org/10.1007/978-3-030-85082-1_16 - PDF: https://www.gyunam.com/publication/2021-simpda/paper.pdf - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) - Tags: Conference Proceedings, Object-Centric Process Mining, SIMPDA2021 ## Abstract The application of process mining techniques to real-life information systems is often challenging. Considering a Purchase to Pay (P2P) process, several case notions such as order and item are involved, interacting with each other. Therefore, creating an event log where events need to relate to a single case (i.e., process instance) leads to convergence (i.e., the duplication of an event related to different cases) and divergence (i.e., the inability to separate events within the same case) problems. To avoid such problems, object-centric event logs have been proposed, where each event can be related to different objects. These can be exploited by a new set of process mining techniques. This paper describes OCEL (Object-Centric Event Log), a generic and scalable format for the storage of object-centric event logs. The implementation of the format can use either JSON or XML, and tool support is provided. ## Practical problem Companies that want to analyze processes such as purchase-to-pay or order-to-cash from their ERP systems find that each event involves several interacting business objects, such as orders, items, packages and customers, rather than one case. Forcing such data into the standard single-case event log format (XES) duplicates events across cases (convergence) or blurs the order of events inside a case (divergence), distorting discovered models and performance figures. Earlier object-centric formats such as XOC replicate object attributes in every event and scale poorly with the number of events and objects, while tabular files lose typing. Analysts and tool builders need a common, compact, strongly typed exchange format for event data with multiple object types. ## Research questions 1. How should object-centric event logs be structured so that events relate to multiple typed objects without the convergence and divergence problems of single-case logs? 2. How can such a format be serialized and supported by tools so that it scales better than existing proposals such as XOC? ## Contributions - The OCEL (Object-Centric Event Log) standard: a meta-model in which a log holds events and objects, each event references a set of objects of different types, and object attributes are stored once rather than per event. - Support for strongly typed and nested elements (lists and maps) and global default elements for logs, events and objects. - Serializations in JSON and XML with validation schemas, specified formally at ocel-standard.org. - Tool support in Python (the ocel-support library on GitHub) and Java (`ProM` OCELStandard package) for import, export, schema validation and flattening to traditional event logs. - Example object-centric event logs extracted from the SAP IDES system in both serializations. ## Key findings - The paper is a standard proposal without experiments; it argues that decoupling object attributes from events avoids the replication that makes XOC logs grow badly with the number of events and objects. - It demonstrates the format with a JSON-OCEL example of an order handling process with customer, item, order and package object types, and shows import, export and validation in Python and Java. - Open limitations named by the paper: no consistency checks for advanced constraints such as the number of objects per event, no support for event streams, whole-log in-memory loading, and no real-world case studies yet. --- # Quality-Aware Resource Model Discovery - Type: Journal article - URL: https://www.gyunam.com/publication/2021-applied-sciences/ - Date: 2021-06-21 - Authors: Minsu Cho, Gyunam Park, Minseok Song, Jinyoun Lee, Euiseok Kum - Venue: Applied Sciences - DOI: https://doi.org/10.3390/app11125730 - PDF: https://www.gyunam.com/publication/2021-applied-sciences/paper.pdf - Research lines: [Process discovery and diagnostics](https://www.gyunam.com/research-line/process-discovery-and-diagnostics/index.md) - Tags: Journal Article, Prediction, DSS ## Abstract Context-aware process mining aims at extending a contemporary approach with process contexts for realistic process modeling. Regarding this discipline, there have been several attempts to combine process discovery and predictive process modeling and context information, e.g., time and cost. The focus of this paper is to develop a new method for deriving a quality-aware resource model. It first generates a resource-oriented transition system and identifies the quality-based superior and inferior cases. The quality-aware resource model is constructed by integrating these two results, and we also propose a model simplification method based on statistical analyses for better resource model visualization. This paper includes tooling support for our method, and one of the case studies on a semiconductor manufacturing process is presented to validate the usefulness of the proposed approach. We expect our work is practically applicable to a range of fields, including manufacturing and healthcare systems. ## Practical problem In semiconductor fabs, hospitals and service operations the quality of the outcome depends heavily on which machines or staff handle each step. Quality engineers know this but have no systematic way to see which paths through hundreds of resources lead to good or bad results. Process mining tools describe control flow and sometimes time or cost, but quality is rarely treated as context and resource networks are not linked to it. When the resource network is drawn for a long process with many machines per step, it has thousands of nodes and arcs and cannot be read, so the insight stays hidden in the data. ## Research questions 1. Can superior and inferior resource paths be identified from real-life manufacturing event logs annotated with case quality? 2. Does simplifying the resource model based on the statistical effect of each resource on quality make the model easier to understand? ## Contributions - A five-step method for discovering quality-aware resource models: resource-oriented transition system discovery, superior and inferior case identification, integrated model construction, statistical analysis of resources, and model simplification. - Formalizes case-annotated transition systems and the superior-case ratio, inferior-case ratio, between-sum and between-difference measures used to rank and color states and arcs. - An algorithm that classifies cases as superior, inferior or moderate using a Kolmogorov-Smirnov normality test followed by mean and standard deviation or median and median absolute deviation bands. - A statistical procedure per activity (ANOVA with linear contrasts, or Kruskal-Wallis with Mann-Whitney U tests) that labels resources as superior, inferior or moderate, with three merging rules for simplification. - A Python tool with a graphical user interface covering import, preprocessing, model mining, statistical analysis, interactive filtering and highlighting, and export. ## Key findings - On a Korean semiconductor manufacturing log of about 15,000 wafers and 118 analyzed activities (selected from more than 500 with domain experts), a weight of 0.85 separates 20 percent superior and 20 percent inferior cases and yields an integrated model with 120 vertical activity lines. - Simplification with all significance thresholds at 0.05 and merging type 1 reduces the model from 1,773 to 504 nodes (71.6 percent) and from 11,973 to 1,078 arcs (91 percent). - In the most complex activity (the 51st line) the number of states drops from 93 to 12, with 6 resources flagged as producing statistically significant high quality and 5 as producing low quality. - Limitation: the method still needs user-set parameters (weights, thresholds) and only explains past resource-quality relations; it does not predict the quality of a running case or operate online. --- # A Scalable Database for the Storage of Object-Centric Event Logs - Type: Journal article - URL: https://www.gyunam.com/publication/2021-icpm-demo/ - Date: 2021-04-01 - Authors: Alessandro Berti, Anahita Farhang Ghahfarokhi, Gyunam Park, Wil M. P. van der Aalst - Venue: CoRR - PDF: https://www.gyunam.com/publication/2021-icpm-demo/paper.pdf - Link (URL): https://arxiv.org/abs/2202.05639 - Research lines: [Object-centric process mining](https://www.gyunam.com/research-line/object-centric-process-mining/index.md) ## Abstract Object-centric process mining provides a set of techniques for the analysis of event data where events are associated to several objects. To store Object-Centric Event Logs (OCELs), the JSON-OCEL and JSON-XML formats have been recently proposed. However, the proposed implementations of the OCEL are file-based. This means that the entire file needs to be parsed in order to apply process mining techniques, such as the discovery of object-centric process models. In this paper, we propose a database storage for the OCEL format using the MongoDB document database. Since documents in MongoDB are equivalent to JSON objects, the current JSON implementation of the standard could be translated straightforwardly in a series of MongoDB collections. ## Practical problem Companies that want to analyze processes involving several interacting business objects (orders, items, invoices) store their event data in the Object-Centric Event Log (OCEL) standard, which comes as JSON or XML files. Every analysis tool has to read the entire file into memory first, so logs with millions of events exceed the memory of an ordinary workstation and cannot be analyzed at all. Graph databases have been tried as an alternative but do not map directly onto the OCEL specification, and columnar stores handle only basic column types rather than the nested JSON structures OCEL uses. Analysts therefore lack a storage back end that keeps large object-centric logs queryable without loading everything at once. ## Research questions 1. How can the JSON-OCEL format be mapped onto a document database so that object-centric event logs can be stored and queried without loading the whole log into memory? 2. How does such a database back end scale in insertion time, storage size and analysis time as logs grow to hundreds of millions of events? ## Contributions - A MongoDB schema for the OCEL standard that translates the JSON-OCEL events, objects and global sections into document collections. - An indexing scheme (identifier, activity and a multikey index on the object map for events; identifier and type for objects) to speed up queries. - Database-level aggregations that compute object lifecycles, event and object statistics, and object-centric directly-follows graphs without exporting the data. - A Python tool in the `ocel-support` repository (GitHub OCEL-standard/ocel-support) that imports and exports JSON/XML-OCEL to and from MongoDB with example scripts. - A scalability assessment on synthetic logs from 1 million to 100 million events. ## Key findings - On a notebook with an i7-7500U CPU and 16 GB RAM, insertion plus indexing takes 79.64 s for 1 million events and 7,983.87 s for 100 million events. - MongoDB's binary storage roughly halves the footprint (54,642 MB of JSON become 21,304 MB of BSON at 100 million events), but the indexes add a further 16,852 MB. - Computing the object-centric directly-follows graph takes 25.50 s at 1 million events and 3,657 s at 100 million events, with RAM capped at about 6.7 GB while disk usage grows to 36,478 MB. - The in-memory approach ran out of memory at 6.8 million events, whereas the database handled 100 million; the trade-off is markedly slower computation, and the prototype has not been used in a real-life case study. --- # A General Framework for Action-Oriented Process Mining - Type: Conference paper - URL: https://www.gyunam.com/publication/2021-bpm/ - Date: 2021-01-19 - Authors: Gyunam Park, Wil M.P. van der Aalst - Venue: 2020 International Conference on Business Process Management - DOI: https://doi.org/10.1007/978-3-030-66498-5_16 - PDF: https://www.gyunam.com/publication/2021-bpm/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) - Tags: Conference Proceedings, Action-Oriented Process Mining, BPM2020 ## Abstract Process mining provides techniques to extract process-centric knowledge from event data available in information systems. These techniques have been successfully adopted to solve process-related problems in diverse industries. In recent years, the attention of the process mining discipline has shifted to supporting continuous process management and actual process improvement. To this end, techniques for operational support, including predictive process monitoring, have been actively studied to monitor and influence running cases. However, the conversion from insightful diagnostics to actual actions is still left to the user (i.e., the “action part” is missing and outside the scope of today’s process mining tools). In this paper, we propose a general framework for *action-oriented process mining* that supports the continuous management of operational processes and the automated execution of actions to improve the process. As proof of concept, the framework is implemented in *ProM*. ## Practical problem Operations managers use process mining to spot bottlenecks, rule violations and late orders, but today's tools stop at diagnostics: a report, a dashboard or a prediction. Turning those insights into concrete actions, such as reprioritizing an order, assigning more staff or alerting a case manager, is left entirely to people, and a one-off report quickly goes stale as conditions change. Operational support techniques can monitor running cases continuously, yet they do not say how the results should be acted upon, and commercial action engines do not handle streaming data or process-level interventions. The step from insight to automated improvement is missing. ## Research questions 1. How can process mining diagnostics on a continuously updated event stream be converted automatically into management actions executed by the information system? 2. Does a constraint monitor detect violations in such a stream, and does an action engine generate the corresponding transactions? 3. Does applying the generated transactions improve the operational process? ## Contributions - A general framework for action-oriented process mining with two components: a constraint monitor that turns an event stream into a constraint instance stream, and an action engine that turns it into an action instance stream of transactions. - Formalizes event streams over multiple object classes (a conventional event log being a special case), time windows, time moments, contexts, constraint formulas, constraints, transactions, action formulas and their instances. - Distinguishes three kinds of actions: process-level valves, instance-level adaptors and alerts. - A `ProM` plug-in (package ActionOrientedProcessMining) with an XML-based Object-Centric Log format, a Constraint Formula Language, an Action Formula Language and an Action Instance Stream format, released on GitHub (gyunamister/ActionOrientedProcessMining). - A proof of concept with a simulated order handling information system and a gateway that translates action instances into system transactions. ## Key findings - In a simulated order handling system with 16 resources, the constraint 'an order must be delivered within 72 hours' is evaluated every 24 hours, and every one of 40 inspected orders whose delivery time exceeds 72 hours is detected by the constraint monitor. - The action engine generates the priority-raising transaction for every violated order and the case manager notification whenever a violation lasts longer than 24 hours. - Over a 30-day simulation, the number of violated instances is lower when the generated transactions are applied than when they are not. - The evaluation is a proof of concept on an artificial system without quantified effect sizes; validation on real systems such as SAP or Salesforce and a taxonomy of constraints and actions are left as future work. --- # From predictive to prescriptive process monitoring: Recommending the next best actions instead of calculating the next most likely events - Type: Conference paper - URL: https://www.gyunam.com/publication/2020-wi/ - Date: 2020-03-01 - Authors: Sven Weinzierl, Sandra Zilker, Matthias Stierle, Gyunam Park, Martin Matzner - Venue: 14th Internationale Tagung Wirtschaftsinformatik - DOI: https://doi.org/10.30844/wi_2020_c12-weinzierl - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) - Tags: Conference Proceedings, Recommendation, WI2020 ## Abstract Predictive business process monitoring (PBPM) deals with predicting a process's future behavior based on historical event logs to support a process's execution. Many of the recent techniques utilize a machine-learned model to predict which event type is the next most likely. Beyond PBPM, prescriptive BPM aims at finding optimal actions based on considering relevant key performance indicators. Existing techniques are geared towards the outcome prediction and deal with alarms for interventions or interventions that do not represent process events. In this paper, we argue that the next event prediction is insufficient for practitioners. Accordingly, this research-in-progress paper proposes a technique for determining next best actions that represent process events. We conducted an intermediate evaluation to test the usefulness and the quality of our technique compared to the most frequently cited technique for predicting next events. The results show a higher usefulness for process participants than a next most likely event. --- # Predicting performances in business processes using deep neural networks - Type: Journal article - URL: https://www.gyunam.com/publication/2020-dss/ - Date: 2020-01-02 - Authors: Gyunam Park, Minseok Song - Venue: Decision Support Systems - DOI: https://doi.org/10.1016/j.dss.2019.113191 - PDF: https://www.gyunam.com/publication/2020-dss/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) - Tags: Journal Article, Prediction, DSS ## Abstract Online operational support is gaining increasing interest due to the availability of real-time data and sufficient computing power, such as predictive business process monitoring. Predictive business process monitoring aims at providing timely information that enables proactive and corrective actions to improve process enactments and mitigate risks. There are a handful of research works focusing on the predictions at the instance level. However, it is more practical to predict the performance of processes at the process model level and detect potential weaknesses in the process to facilitate the proactive actions that will improve the process execution. Thus, in this paper, we propose a novel method to predict the future performances of a business process at the process model level. More in detail, we construct an annotated transition system and generate a process representation matrix from it. Based on the process representation matrix, we build performance prediction models using deep neural networks that consider both spatial and temporal dependencies present in the underlying business process. To validate the proposed method, we performed case studies on three real-life logs. ## Practical problem Operations managers in hospitals, banks, and service desks want to know where their process will clog up next so they can add staff or reroute work before delays build. Predictive monitoring tools exist, but they forecast one running case at a time (for example the remaining time of a single loan application), and a manager of a process with thousands of concurrent cases cannot act on thousands of individual alerts. What is missing is a forecast at the level of the process model itself: which activities and hand-overs will have long waiting or processing times in the next hour or next day. Simple per-activity statistics ignore the fact that a delay in one step propagates to the next, so they miss exactly the cascading bottlenecks managers care about. ## Research questions 1. How can the future performance of a business process be predicted at the process model level, rather than per instance, from historical event data? 2. Do deep neural networks that capture both the spatial structure of the process model and its temporal evolution predict activity and transition performance more accurately than per-location statistical models or similarity search? ## Contributions - Formalizes an annotated transition system in which states and transitions carry multisets of waiting, processing, and sojourn time measurements obtained by replaying the event log. - Two process representation matrices: a two-dimensional matrix (model elements by time windows) and a three-dimensional matrix (from-state by to-state by time window) encoding directly-follows topology. - Three prediction models adapted from traffic congestion forecasting: a convolutional neural network (CNN), a long short-term memory network (LSTM), and a long-term recurrent convolutional network (LRCN). - Two baseline families for model-level performance prediction: per-element statistical regressors (linear regression, random forest, support vector regression) and nearest-snapshot search with Euclidean, Chebyshev, and cosine distance. - A Python implementation with source code and supplementary material released on GitHub (`gyunamister/performance_prediction`), evaluated on three real-life logs. ## Key findings - On an emergency-department log (459,700 events, 29,871 patients) LRCN gives the lowest error in all four tasks; for 1-hour-ahead state sojourn time its mean absolute error is 0.26 h versus 0.36 h for the best baseline (support vector regression). - On BPI Challenge 2012 (262,200 events, 13,087 cases), predicting 6-hour-ahead processing time from 144 hours of history, CNN and LRCN reach 0.44 h state and 0.19 h transition mean absolute error, beating all statistical and search-based baselines. - On the Helpdesk log (8,988 events, 2,542 cases), predicting 5 days ahead from 30 days of history, LRCN attains the lowest state error (1.71 days versus 2.24 days for support vector regression). - Encoding the directly-follows topology in the three-dimensional matrix is what makes LRCN stable: as the transition-system horizon grows on BPI Challenge 2012, LRCN improves most, at the cost of longer training and higher overfitting risk. - Deep models do not win everywhere: LSTM falls behind baselines for short-term prediction from only 12 hours of history, CNN and LSTM trail statistical models for 3-hour-ahead transition prediction, and Euclidean search is best for Helpdesk transitions. --- # Prediction-based Resource Allocation using LSTM and Minimum Cost and Maximum Flow Algorithm - Type: Conference paper - URL: https://www.gyunam.com/publication/2019-icpm/ - Date: 2019-06-26 - Authors: Gyunam Park, Minseok Song - Venue: 2019 International Conference on Process Mining - DOI: https://doi.org/10.1109/ICPM.2019.00027 - PDF: https://www.gyunam.com/publication/2019-icpm/paper.pdf - Research lines: [Prescriptive process analytics and digital twins](https://www.gyunam.com/research-line/prescriptive-process-analytics/index.md) - Tags: Conference Proceedings, Recommendation, ICPM2019 ## Abstract Predictive business process monitoring aims at providing the predictions about running instances by analyzing logs of completed cases of a business process. Recently, a lot of research focuses on increasing productivity and efficiency in a business process by forecasting potential problems during its executions. However, most of the studies lack suggesting concrete actions to improve the process. They leave it up to the subjective judgment of a user. In this paper, we propose a novel method to connect the results from predictive business process monitoring to actual business process improvements. More in detail, we optimize the resource allocation in a non-clairvoyant online environment, where we have limited information required for scheduling, by exploiting the predictions. The proposed method integrates offline prediction model construction that predicts the processing time and the next activity of an ongoing instance using LSTM with online resource allocation that is extended from the minimum cost and maximum flow algorithm. To validate the proposed method, we performed experiments using an artificial event log and a real-life event log from a global financial organization. ## Practical problem Hospitals, banks and other service organizations must decide many times a day which staff member takes which pending task, and a poor choice lengthens waiting for everyone, especially urgent cases. Scheduling heuristics from manufacturing assume the processing time and the next step of every job are known in advance, but in an emergency department or a loan desk nobody knows how long the current step will take or what the next one will be until it happens. Predictive monitoring tools can forecast such quantities, yet they stop at the forecast and leave the decision to the manager's judgment. The gap between a prediction and a concrete allocation decision therefore remains open in practice. ## Research questions 1. How can prediction models supply the scheduling parameters, namely processing time and next activity, that are unknown in a non-clairvoyant online environment? 2. How can resources be dispatched efficiently on the basis of these predictions so that total weighted completion time is minimized? ## Contributions - Formulates resource allocation in business processes as a non-clairvoyant online job shop scheduling problem that minimizes total weighted completion time under limited information. - A two-phase method: offline training of a multi-task Long Short-Term Memory (LSTM) network that predicts processing time and next activity from one-hot encoded prefixes, and online scheduling driven by an event stream. - A scheduling algorithm that builds a bipartite graph of ready and predicted work items versus resources, with edge cost (processing time plus readiness penalty) divided by instance weight, and solves minimum cost maximum flow with the network simplex method. - Pseudo-assignments that reserve resources for predicted high-weight work items before they are released, followed by execution of the matches whose instance and resource are both available. - An evaluation against a WeightGreedy first-come-first-served baseline on a simulated emergency department log and the BPI Challenge 2012 loan application log. ## Key findings - On the simulated emergency department (11 activities, 25 resources, 1,000 training patients, stream of 100 patients) the method beats WeightGreedy for every number of instances; at 80 instances the baseline total weighted completion time of 28,393 is 14 percent above the method's 24,804. - On one day of manual work items from BPI Challenge 2012 (110 instances, about 3 activities each, weights from 10 requested-amount buckets) total weighted completion time falls from 1,479 for the baseline to 1,038, i.e. the baseline is 42 percent higher. - The gain costs computation: 115.6 seconds versus 7.6 for the baseline on the real log, with 110.1 seconds spent on predictions because every work item has many candidate resources. - Limitation: the method depends on prediction quality; a wrong processing-time prediction assigns the work item to a suboptimal resource, and repeated prediction for every newly released instance dominates runtime. --- ## Teaching (12) --- # Foundations of Data Analytics - Type: Course - URL: https://www.gyunam.com/teaching/foundations-data-analytics/ - Date: 2026-01-01 - Period: 2026 – present - Role: Lecturer - Course type: Undergraduate course - Institution: Eindhoven University of Technology --- # Foundation Data Challenge - Type: Course - URL: https://www.gyunam.com/teaching/foundation-data-challenge/ - Date: 2026-01-01 - Period: 2026 – present - Role: Lecturer - Course type: Undergraduate course - Institution: Eindhoven University of Technology --- # Seminar Process Analytics - Type: Course - URL: https://www.gyunam.com/teaching/pa-seminar/ - Date: 2025-09-01 - Period: 2025 – present - Role: Lecturer - Course type: Graduate course - Institution: Eindhoven University of Technology --- # Process Discovery Using Python - Type: Course - URL: https://www.gyunam.com/teaching/software-lab/ - Date: 2022-04-01 - Period: 2022 – 2025 - Role: Lecturer - Course type: Graduate lab course - Institution: RWTH Aachen University --- # Introduction to Process Discovery Using Python - Type: Course - URL: https://www.gyunam.com/teaching/intro-process-discovery-python/ - Date: 2022-04-01 - Period: 2022 – 2025 - Role: Lecturer - Course type: Undergraduate lab course - Institution: RWTH Aachen University --- # Introduction to Data Science - Type: Course - URL: https://www.gyunam.com/teaching/ids/ - Date: 2019-10-01 - Period: 2019 – 2022 - Role: Instructor - Course type: Graduate course - Institution: RWTH Aachen University --- # Selected Topics in Process Mining - Type: Course - URL: https://www.gyunam.com/teaching/seminar-selected-topics/ - Date: 2021-04-01 - Period: 2021 - Role: Instructor - Course type: Graduate course - Institution: RWTH Aachen University --- # Business Process Intelligence - Type: Course - URL: https://www.gyunam.com/teaching/bpi/ - Date: 2020-04-01 - Period: 2020 - Role: Instructor - Course type: Undergraduate course - Institution: RWTH Aachen University --- # AI and IoT Technology Training Program - Type: Course - URL: https://www.gyunam.com/teaching/ai-iot-training/ - Date: 2018-01-01 - Period: 2018 - Role: Tutor - Course type: Training program - Institution: POSTECH --- # POSCO AI-Expert Training Program - Type: Course - URL: https://www.gyunam.com/teaching/posco-ai-expert/ - Date: 2018-01-01 - Period: 2018 - Role: Tutor - Course type: Training program - Institution: POSCO --- # Introduction to Optimization - Type: Course - URL: https://www.gyunam.com/teaching/intro-optimization/ - Date: 2018-01-01 - Period: 2018 - Role: Teaching Assistant - Course type: Undergraduate course - Institution: POSTECH --- # Database System - Type: Course - URL: https://www.gyunam.com/teaching/database-system/ - Date: 2017-01-01 - Period: 2017 - Role: Teaching Assistant - Course type: Undergraduate course - Institution: POSTECH