ClaimGraph: Relational Detection of Coordinated Anomalies in Insurance Claims

Abstract

Insurance fraud frequently involves coordinated behavior across claimants, service providers, vehicles, and payment accounts, making isolated claim-level screening insufficient. This paper presents ClaimGraph, a relational anomaly detection approach that models heterogeneous interactions among entities appearing in insurance claims. Rather than assigning risk scores solely from claim attributes, the method captures recurring entity combinations, unusually dense local structures, and temporal patterns of shared activity. Experiments are conducted on 1.26 million claims involving 2.8 million entities over a three-year period, including 18,460 claims confirmed as fraudulent through subsequent investigation. ClaimGraph achieves an AUPRC of 0.612 and AUROC of 0.947, compared with 0.438 and 0.901 for a gradient-boosted claim-level classifier. At a fixed investigation capacity of the highest-risk 1% of claims, the proposed method identifies 42.7% of confirmed fraud cases, representing a 13.4-percentage-point improvement over the strongest baseline. Further analysis indicates that shared payment accounts and repeated claimant–provider relationships are particularly informative for detecting coordinated fraud. The results demonstrate that relational structure can reveal suspicious activity that remains difficult to identify from individual claims alone.