XG-ScamNet: An Explainable Graph Neural Network Framework for Adaptive Financial Scam Identification and Prevention

Abstract

Financial scams increasingly exploit the relational structure of digital payment networks, making transaction level anomaly detection insufficient on its own. This paper proposes an adaptive graph neural network framework that models financial transactions as a dynamic graph of accounts and transfers, learns time aware representations through an attention-based encoder with adaptive edge reweighting, and produces scam predictions accompanied by human readable explanations. The framework combines a graph attention encode, a temporal memory update mechanism that tracks behavioural drift, and a post hoc explainability module that generates subgraph and feature level attributions for every flagged account. Experiments on a large-scale synthetic and semi real transaction dataset show that the proposed model achieves a precision of 0.94, a recall of 0.92, and an F1 score of 0.93, outperforming logistic regression, random forest, graph convolutional network, and standard graph attention network baselines. An ablation study confirms that adaptive edge reweighting and temporal memory each contribute measurable gains, and a fidelity evaluation shows that the explanations generated by the framework align closely with the features actually used by the model, consistent with fidelity criteria proposed in prior explainability literature. The results indicate that adaptive graph learning paired with explainability can deliver both higher detection accuracy and greater analyst trust, which are the two properties most needed for deployment in real financial institutions.