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

A team of researchers has published a paper proposing a novel neuro-symbolic AI framework designed to solve a critical flaw in current predictive process monitoring systems. Traditional AI models are purely sub-symbolic, learning patterns from data but often ignoring hard business rules and compliance constraints. This leads to impractical or non-compliant predictions, such as scheduling a surgery before a mandatory waiting period has elapsed. The new approach, led by Fabrizio De Santis, Gyunam Park, and Wil van der Aalst, formally integrates symbolic logic with neural networks to create a 'compliance-aware' predictor.

The core innovation is the use of Logic Tensor Networks (LTNs), a neuro-symbolic architecture, to inject explicit process knowledge directly into the model's learning objective. The method follows a four-stage pipeline: extracting features from event logs, extracting domain rules (like regulatory constraints), creating a formal knowledge base, and then injecting this knowledge during model training. This forces the AI to respect real-world logic while still learning from data.

Evaluation results demonstrate a dual benefit: the model not only learns to adhere to the injected constraints but also achieves better overall predictive performance compared to standard baseline models. This indicates that incorporating domain knowledge doesn't just make predictions compliant—it makes them more accurate by guiding the model away from nonsensical data correlations. The work was accepted at the prestigious CAiSE 2026 conference.