Abstract
Financial fraud remains a persistent and evolving threat to the stability of global economic systems, necessitating advanced detection mechanisms that can adapt to sophisticated criminal strategies. Traditional supervised learning approaches often struggle with the inherent challenges of fraud detection: extreme class imbalance, label scarcity, and the complex, camouflaged nature of illicit activities. This paper presents a novel framework for detecting fraudulent behavior by modeling financial systems as heterogeneous graphs connecting transactions, accounts, and physical devices. We propose a Contrastive Heterogeneous Graph Neural Network (CHG-Net) that leverages self-supervised contrastive learning to generate robust node embeddings without relying heavily on labeled data. By maximizing the agreement between differently augmented views of the transaction-device network, the model learns to distinguish anomalous structural patterns intrinsic to fraud rings, such as device sharing and high-frequency circular transfers. We evaluate our approach on large-scale real-world financial datasets, demonstrating that CHG-Net significantly outperforms state-of-the-art supervised Graph Neural Networks (GNNs) and traditional machine learning baselines in terms of AUC and Recall, particularly in scenarios with limited labeled training data.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Copyright (c) 2025 Jie Hu (Author)