Fraud Detection and Risk Assessment of Online Payment Transactions on E-Commerce Platforms Based on LLM and GCN Frameworks

With the rapid expansion of e-commerce, online payment fraud has become increasingly sophisticated, posing significant challenges to financial security and undermining consumer trust. Traditional detection methods often struggle to capture the intricate relational and behavioral patterns inherent in transactional data, limiting their effectiveness in highly imbalanced scenarios. This study introduces a novel fraud detection framework that integrates Large Language Models (LLMs) with Graph Convolutional Networks (GCNs) to enhance the identification of fraudulent activities in online payment systems. A dataset comprising 2,840,000 transactions was collected over a 14-day period from major e-commerce platforms, including Amazon, involving approximately 2,000 U.S.-based consumers and 30 merchants. Among these transactions, fewer than 6,000 were fraudulent, presenting a highly skewed class distribution that reflects real-world conditions. In this framework, consumers and merchants were represented as nodes, and transactions as edges, forming a heterogeneous graph that captures both direct and indirect interactions. A GCN was applied to this graph to learn complex structural and relational patterns, effectively modeling the interdependencies among participants. In parallel, semantic features were extracted from transaction metadata using GPT-4o and Tabformer, capturing textual and categorical information that complements the structural graph features. By fusing these semantic representations with graph-based features, the model can identify subtle and context-dependent indicators of fraudulent behavior that traditional methods may overlook. Experimental results demonstrate that the proposed framework achieves an overall accuracy of 0.98, with a balanced performance in terms of precision and sensitivity, highlighting its robustness in detecting rare fraudulent instances within a massive dataset. The integration of LLM-derived semantic features with graph structural information significantly improves detection efficacy compared with models relying solely on either approach. This hybrid approach offers a scalable, real-time solution for enhancing the security of online payment environments and provides a promising avenue for applying graph-based deep learning techniques in financial fraud prevention.