Explainable Graph-Based Financial Fraud Detection Using Machine Learning and Data Analytics

Financial fraud constitutes one of the most economically destructive threats facing the global financial ecosystem, causing estimated annual losses exceeding USD 5.1 trillion across banking, insurance, e-commerce, and fintech sectors. Conventional rule-based fraud detection frameworks suffer from elevated false-positive rates, an inability to model complex relational transaction patterns across accounts and merchants, and complete opacity in their decision rationale — rendering regulatory compliance and analyst trust difficult to sustain. This paper presents XGFFD (Explainable Graph-Based Financial Fraud Detection), a unified system integrating Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN) with an XGBoost-led ensemble classifier comprising LightGBM and Random Forest to detect fraudulent transactions by modelling financial entities as heterogeneous graphs. SHAP (SHapley Additive exPlanations) and GNNExplainer are incorporated for post-hoc model interpretability, producing human-readable feature attribution and subgraph visualisation for every fraud prediction. Evaluated on the IEEE-CIS Fraud Detection dataset and a synthetic heterogeneous transaction graph with 2.1 million nodes and 8.7 million edges, XGFFD achieved an F1-score of 0.9312, a Precision-Recall AUC of 0.9478, and a Matthews Correlation Coefficient (MCC) of 0.8941, surpassing all tabular and graph-only baselines. The explainability layer reduced analyst investigation time by 63%, and the system sustained 12,000 transactions per second on a four-GPU cluster, confirming production-grade viability.