# 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 Pure rule-following is too rigid, because the rules live in text and in people's heads, some are outdated or contradicted by practice, and they say nothing about what is likely to happen next.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/)).