Machine Learning-Driven Framework for Financial Fraud Detection Using Graph Neural Networks and Explainable AI

Abstract

Financial fraud causes global losses exceeding USD 5.1 trillion annually, and the rapid digitization of financial services continues to expand opportunities for fraudulent activities. Traditional rule-based and machine learning fraud detection systems are limited because they analyze transactions in isolation and struggle to adapt to evolving fraud patterns. To address these challenges, this paper presents ML-GNN, a Machine Learning-Driven Graph Neural Network framework for multi-class financial fraud detection that models financial ecosystems as heterogeneous and dynamically evolving transaction graphs. The framework combines a multi-source data preprocessing pipeline with SMOTE-based class balancing and SHAP-guided feature selection, a graph construction module representing accounts, merchants, devices, and IP addresses as interconnected nodes, a hybrid Graph Neural Network integrating GraphSAGE, Graph Attention Networks (GAT), and Temporal GNN with LSTM gating, and a stacking ensemble that merges graph-based embeddings with XGBoost predictions through a logistic regression meta-learner. Experimental evaluation on the IEEE-CIS Fraud Detection, PaySim Synthetic Financial Transactions, and Elliptic Bitcoin Transactions datasets, comprising more than 7.15 million transactions, demonstrates that ML-GNN achieves an AUC-ROC of 0.987, a weighted F1-score of 95.6%, and a false positive rate of 2.1%, outperforming the strongest standalone GNN baseline by 5.6 percentage points in AUC and 4.5 points in F1-score. Furthermore, SHAP and GNNExplainer analyses show that the model’s predictions are based on interpretable graph-structural features, supporting transparency and regulatory compliance requirements such as GDPR Article 22.