Robust Explainable Supplier Risk Prediction under Imperfect Domain Knowledge: Empirically Vetted Monotonic XGBoost with TreeSHAP
TL;DR
The study shows that domain constraints should be auditable inputs rather than unquestioned rules and proposes empirically vetted monotonic XGBoost (EV-Mono), in which domain experts still supply candidate directions while training data are allowed only to veto, never reverse, a candidate constraint.
Abstract
Monotonic constraints are attractive for supplier-risk models because they can force predictions to respect domain knowledge, but a wrong expert sign is then enforced everywhere. This paper studies that overlooked failure mode and proposes empirically vetted monotonic XGBoost (EV-Mono), in which domain experts still supply candidate directions while training data are allowed only to veto, never reverse, a candidate constraint. A training-only bootstrap partial-dependence gate is combined with exact TreeSHAP, a prior-misspecification stress test (PMST), and a new Explanation Intervention Consistency (XIC) metric that checks whether risk-reducing feature interventions move the corresponding attribution in the expected direction. Controlled experiments use 12,000 fully synthetic supplier records with 18 known monotone risk drivers, six null variables, nonlinear main effects and sign-preserving interactions; an adverse covariate-shift test contains 3,000 additional records. With correct priors, EV-Mono attains AUC 0.947, zero monotonicity violations, XIC 1.000, and places all 18 informative features in the top 18 TreeSHAP ranking. Under 40% deliberately corrupted priors, hard monotonic XGBoost falls to AUC 0.932, explanation-coherence rate 0.611 and monotonicity-violation rate 0.343; EV-Mono rejects all seven wrong signs and retains AUC 0.946, coherence 1.000 and violation rate 0.091. The study shows that domain constraints should be auditable inputs rather than unquestioned rules.