Abstract
Given a running process case, prescriptive process analytics (PrPA) recommends the next activity that best improves a desired outcome, such as reducing cycle time. Recommendations must also be operationally feasible: an activity that improves the predicted outcome but violates process constraints cannot be enacted. The core challenge is therefore not compliance alone, which inference-time filtering can enforce, but achieving high outcome quality within the compliant space. Existing approaches fall short: post-hoc filtering never learns which compliant activities are outcome-improving, while logic-as-loss penalties balance unstable competing gradients from data fit and compliance. Both also treat all constraints equally, so combining mandatory hard constraints with relaxable soft constraints can leave many cases with no feasible recommendation. We propose NeSyPA, a neurosymbolic PrPA method with two contributions: Hierarchical Constraint Relaxation, which drops soft constraints only when the full constraint set is unsatisfiable, and Constraint-Compiled Outcome Distillation, which compiles per-prefix feasible sets into feasibility-masked, outcome-weighted soft targets that teach the neural prescriptor to rank feasible activities by outcome merit. On four real-world event logs, NeSyPA achieves the highest Effective Compliance Rate on all test sets, with gains of +6.7% to +9.9%, driven by higher outcome quality within the compliant space.
Publication
Business Process Management
Practical problem
Case workers handling a running loan application or a traffic fine want a system that tells them which step to take next so the case ends well, for instance with an accepted offer or a fine paid on time. Such a recommendation is useless if it breaks the rules of the process: skipping a mandatory review may look good for cycle time but can never be enacted. Today’s recommenders either train without rules and filter out forbidden steps afterwards, so they never learn which allowed steps actually help, or they add a rule-violation penalty to training, which forces an unstable balancing act between fitting data and obeying rules. On top of that, real rule books mix regulations that must never be broken with habitual patterns mined from history, and when both are enforced equally, a large share of cases ends up with no permitted next step at all.
Research questions
- How can a neural recommender be trained to rank only constraint-compliant next activities by their expected outcome improvement, without a tunable trade-off between compliance and outcome quality?
- How should mandatory (hard) and relaxable (soft) process constraints be prioritized so that cases with over-constrained knowledge bases still receive a feasible recommendation?
Contributions
- NeSyPA, a neurosymbolic prescriptive process analytics method combining a Transformer prescriptor with Declare constraints compiled into deterministic finite automata.
- Hierarchical Constraint Relaxation (HCR), which partitions the knowledge base into hard and soft constraints and drops the soft tier only when the full set admits no activity.
- Constraint-Compiled Outcome Distillation (CCOD), which turns per-prefix feasible sets into feasibility-masked, outcome-weighted soft targets and fine-tunes the activity head in one pass.
- Effective Compliance Rate (ECR), a metric counting recommendations that are both outcome-improving and constraint-compliant, alongside outcome-improvement and constraint-satisfaction rates.
- An evaluation protocol with four baselines (unconstrained, post-hoc filtering, reinforcement learning, logic-as-loss) on four real-life test sets, with code and data on GitHub (gyunamister/nesy_prpa).
Key findings
- NeSyPA reaches the highest Effective Compliance Rate on all four test sets: 0.569 on BPIC 2012, 0.437 and 0.456 on the two BPIC 2017 splits, 0.399 on Traffic Fines, i.e. +6.7, +9.7, +9.9 and +0.03 percentage points over the best baseline.
- The gains come from outcome quality within the compliant space: Outcome Improvement Rate rises to 0.849, 0.718 and 0.735 on the BPIC logs versus 0.505 to 0.711 for baselines, while constraint satisfaction stays comparable (about 0.60 to 0.68).
- In the ablation, CCOD is the dominant component (+7.1, +11.5, +10.0 and +1.5 points over the post-hoc variant with hierarchy); hierarchy adds at most +1.1 points but cuts irreparable prefixes from 31.8% to 2.3% on BPIC 2012.
- The logic-as-loss baseline degrades constraint satisfaction to 0.753 on Traffic Fines against 0.940 for filtering baselines, and the reinforcement-learning baseline shows high variance (coefficient of variation 0.33 on BPIC 2017 before).
- Injecting 10% spurious soft constraints on BPIC 2012 drops ECR from 0.574 to about 0.345 and then plateaus; relaxation rate and ECR degradation correlate at r = -0.91, and the all-or-nothing soft relaxation remains a limitation.