Neuro-Symbolic Learning for Predictive Process Monitoring via Two-Stage Logic Tensor Networks with Rule Pruning
Summary: arXiv:2603.26944v1 Announce Type: new
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.
Introduction
This article presents a neuro-symbolic approach that integrates domain knowledge as differentiable logical constraints using Logic Tensor Networks (LTNs). By formalizing control-flow, temporal, and payload knowledge through Linear Temporal Logic and first-order logic, the method bridges the gap between traditional data-driven models and the necessity for compliance in various sectors.
Key Contributions
- Two-Stage Optimization Strategy: Our approach introduces a two-stage optimization strategy designed to address the common issue faced by Logic Tensor Networks of satisfying logical formulas while sacrificing predictive accuracy.
- Weighted Axiom Loss: During pretraining, the model employs a weighted axiom loss that prioritizes learning from the data before enforcing logical constraints.
- Rule Pruning: Following the pretraining phase, a rule pruning technique is applied to retain only those axioms that are consistent and contributive based on satisfaction dynamics, enhancing overall model performance.
Evaluation and Results
We conducted evaluations on four real-world event logs to assess the effectiveness of our proposed method. The results demonstrate that the injection of domain knowledge into predictive models significantly enhances performance. More importantly, the two-stage optimization strategy is crucial, as knowledge without proper optimization can lead to substantial degradation in model performance.
Applications
Our neuro-symbolic approach excels in compliance-constrained scenarios, particularly when there are limited compliant training examples. This makes it particularly useful in fields such as:
- Healthcare: Ensuring that medical procedures adhere to established guidelines and sequences.
- Finance: Maintaining compliance with regulatory frameworks in transaction processing.
- Fraud Detection: Identifying anomalous patterns while respecting domain-specific constraints.
Conclusion
The integration of domain knowledge through our neuro-symbolic approach not only enhances predictive accuracy but also ensures adherence to necessary compliance requirements. As industries continue to navigate the complexities of regulatory frameworks, our method provides a robust solution that balances predictive modeling with the constraints imposed by domain-specific rules.
Future Work
Looking ahead, further research will focus on extending this approach to additional domains and exploring the implications of more complex logical constraints. Additionally, we aim to refine the optimization process to further improve performance outcomes in diverse applications.