Neuro-Symbolic Learning for Predictive Process Monitoring via Two-Stage Logic Tensor Networks with Rule Pruning
healthcare
| Source: ArXiv | Original article
A team of researchers from the University of Trento and the Norwegian University of Science and Technology has released a new arXiv pre‑print, “Neuro‑Symbolic Learning for Predictive Process Monitoring via Two‑Stage Logic Tensor Networks with Rule Pruning.” The paper proposes a hybrid architecture that marries deep sequence models with symbolic logic to forecast the next steps in business processes, a capability central to fraud detection, healthcare workflow oversight and supply‑chain risk management.
The core of the method is a two‑stage pipeline. First, a neural encoder—typically a transformer or LSTM—captures temporal patterns in event logs. In the second stage, the encoded representation is fed into a logic tensor network that enforces domain‑specific constraints such as “a payment must follow an invoice” or “a medication dosage cannot exceed a prescribed limit.” A novel rule‑pruning algorithm discards redundant or low‑impact logical clauses, keeping the model both compact and interpretable. Benchmarks on publicly available event‑log datasets (e.g., BPI Challenge 2019 and a hospital admission corpus) show a 5‑7 % lift in prediction accuracy over pure neural baselines while delivering clear explanations for each forecast.
Why it matters is twofold. Accuracy gains translate directly into earlier fraud alerts or timely clinical interventions, reducing financial loss and patient harm. More importantly, the embedded symbolic layer satisfies regulatory demands for traceability: auditors can inspect which business rules drove a prediction, a feature that pure black‑box models lack. The approach also hints at a broader shift toward neuro‑symbolic AI in operational settings, where compliance and explainability are non‑negotiable.
The next steps to watch include a forthcoming evaluation at the International Conference on Business Process Management, where the authors will compare their system against the state‑of‑the‑art diffusion‑based predictors discussed in our March 31 article on A‑SelecT. Industry pilots with Scandinavian banks and a regional health authority are slated for Q3, and the community will be keen to see whether the rule‑pruning technique scales to the massive, noisy logs typical of real‑world deployments.
Sources
Back to AIPULSEN