CGGS: Confidence-Guided Graph Sampling for Scalable GNN Fraud Detection on Imbalanced Financial Networks
Abstract
Graph Neural Networks (GNNs) offer a powerful paradigm for financial fraud detection by exploiting the relational structure of transaction networks to identify suspicious behaviour and coordinated fraud. Translating them into a viable production system, however, remains hard under strict memory, compute and latency budget – costs that span the entire model lifecycle, not merely GNN training. A real transaction graph of hundreds of millions of accounts is overwhelmingly a graph of benign nodes, so much of that lifecycle budget is spent processing nodes that carry little discriminative signal. The same benign majority drives a second challenge, fraud-class homophily collapse: because fraudsters deliberately interact with legitimate entities to camouflage and evade detection, the proportion of same-class edges incident on fraud nodes drops and the fraud class is driven into high heterophily, so neighbourhood aggregation averages the fraud signal away while imbalance biases the classifier toward the benign majority. We propose CGGS (Confidence-Guided Graph Sampling), a two-stage framework that turns both liabilities into a single opportunity at no additional compute cost. Stage 1 trains a lightweight tabular scorer – gradient-boosted trees (XGBoost) in our instantiation – on the node features, ranking every candidate entity by fraud probability without touching graph topology. Stage 2 prunes the confidently-benign entities before the graph is assembled, so the GNN trains only on the compact, higher-homophily subgraph of uncertain, high-risk, and informative nodes. We prove that confidence-guided pruning increases the expected fraud-class homophily and reduces the effective imbalance ratio under a mild ranking assumption. Experiments on a production-scale proprietary transaction graph and two public benchmarks show that CGGS reduces graph size and lifecycle computational cost while outperforming full-graph training and alternative pruning strategies, with ablations confirming robustness across pruning thresholds.