A heterogeneous graph representation learning approach for fraud detection in supply chain networks

In the current cybersecurity era, fraudulent activities in Supply Chain Networks (SCN) have become increasingly sophisticated and can have a huge financial and reputational impact on enterprises’ business. Traditional tabular and sequential models, struggle to capture the network structure of SCN relationships and inter-dependencies. To this end, a novel graph representation learning and deep Graph Neural Network(GNN)–based model for fraud detection in SCN is proposed in this study. In doing so, a heterogeneous graph representation of SCN, where transactions, orders, products, and customers are modeled as interconnected nodes with rich attributes. Then, Graph Autoencoder (GAE) is used to learn low-dimensional node embeddings that preserve both topological structure and attribute information, while Node2Vec captures complementary higher-order structural patterns through biased random walks. These embeddings are integrated within deep GNN variants such as GraphSAGE and GAT. Moreover, extensive experiments conducted on a real-world SCN dataset demonstrate that the proposed model achieves superior detection performance compared to traditional machine learning baselines, in terms of F1 score = 0.904, recall = 0.956, and accuracy = 0.979. The results highlight the effectiveness of combining representation learning with deep GNN variants in capturing hidden relational patterns, thereby providing a scalable and robust solution for fraud detection in complex SCN.