Abstract
Graph neural networks detect organized payment fraud by exploiting shared cards, devices and e-mail accounts, yet deployed detectors face heterogeneous drift that erodes not only accuracy but also the calibration of risk scores and the stability of explanations. This paper presents a drift-resilient explainable graph-learning framework that evaluates and maintains these three properties jointly. A GraphSAGE detector is equipped with exact gradient×input attributions computed by forward-mode differentiation, which separate a transaction's own evidence from evidence propagated through the graph. A label-free monitor standardizes five statistics, namely feature, score, signed prior, decision-focused attribution and structural-reliance drift, against a reference null, and a Diagnose-Then-Act (DTA) policy maps them to the cheapest sufficient action: no action, label-free EM recalibration, delayed-label Platt refresh, or change-point-aware retraining. On FinDrift-G, an IEEE-CIS-style transaction-graph benchmark with ground-truth drift types and fraud mechanisms, each drift type leaves a distinct fingerprint: benign covariate drift drives the feature-PSI z-score to 1073 while PR-AUC stays at 0.909, whereas account-takeover drift reduces the recall of a static GraphSAGE from 0.89 to 0.13. Under harmful drift, DTA reaches a PR-AUC of 0.843 against 0.845 for retraining at every step while using 2.0 instead of 22.3 retraining events per stream, attains the lowest expected calibration error (0.77%), and it retrains only 1.0 times per stream when drift is absent or benign. Discrimination, calibration and explanation stability therefore need to be monitored together, and diagnosing the drift type, rather than merely detecting drift, determines the right maintenance action.