Governance-Aware Entity Resolution: Two-Stage Veto Thresholding for Worst-Case Hard-Negative FPR Control

Punitha Ponnuraj, Sathyan Sugumar

Abstract: Identity Resolution (IR) has an asymmetric risk profile in automated Master Data Management (MDM). While missed mergers often defer record consolidation, false mergers can corrupt downstream systems and compliance. Yet, current model selection relies mostly on aggregate metrics, which often overlook high-cost errors in near-match scenarios. This paper proposes a Two-Stage Veto protocol that decouples model scoring from decision-point selection and enforces a worst-case False Positive Rate (FPR) budget (τ=0.02) on governance-critical non-match sets (hard negatives and high-similarity non-matches). To evaluate this protocol, we developed a deterministic, conflict-driven synthetic benchmark with 6,400 labeled record pairs across nine conflict clusters reflecting enterprise-level issues. Using four fundamental similarity signals (name, address, ZIP consistency, and geospatial proximity), we evaluate Logistic Regression, Random Forest, XGBoost, and three fusion variants. Our results demonstrate that on the blind TEST split at 𝜏 = 0.02, unconstrained thresholding can achieve higher utility but violates the worst-case hard-negative FPR budget by up to an order of magnitude. Under the same constraint, XGBoost remained compliant with the worst-case hard-negative FPR budget, whereas a fusion variant that appeared safe on validation violated the same budget on TEST. External validation on Yelp and the Mannheim (Fodors-Zagats) benchmarks reveal that while model ranking can remain informative in AUC terms across domains, decision thresholds do not, due to score-scale (calibration) shift. These results indicate that IR in production requires explicit governance gates and threshold-survival diagnostics to ensure system integrity.

External IDs:doi:10.36227/techrxiv.176964057.71533616/v1