Compliance-Aware Predictive Process Monitoring: A Neuro-Symbolic Approach

Fabrizio De Santis, Gyunam Park, Wil M. P. Van Der Aalst, Francesco Zanichelli
June, 2026
PDF Cite DOI

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.

Type
Conference paper
Publication
Advanced Information Systems Engineering

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.