Graph Embeddings for Fraud Detection in Mobile Payment Networks: Causal Modeling

Abstract

The rapid expansion of mobile payment networks has revolutionized global financial transactions, providing unprecedented convenience while simultaneously creating complex vulnerabilities exploited by sophisticated malicious actors. Traditional fraud detection systems, which primarily rely on tabular transaction features and isolated machine learning models, frequently fail to capture the intricate, hidden relational patterns among malicious entities. To address this critical limitation, financial technologists have increasingly adopted graph representation learning, utilizing graph embeddings to compress high-dimensional network topologies into dense vectors for predictive modeling. However, the reliance on purely predictive graph embeddings introduces significant vulnerabilities, primarily because these models are highly susceptible to capturing spurious correlations rather than true structural dependencies. This paper investigates the critical integration of causal modeling with graph embeddings to enhance fraud detection accuracy and robustness in mobile payment networks. By applying a causal inference framework to heterogeneous transaction graphs, we distinguish between structural patterns that causally drive fraudulent behavior and those that merely co-occur due to confounding variables. Our comprehensive analysis utilizes a massive, anonymized dataset from a major mobile payment provider, systematically controlling for back-door paths in the network structure. The findings demonstrate that causally informed graph embeddings significantly reduce false positive rates and maintain robust performance under severe concept drift, outperforming state-of-the-art purely correlational baselines. This research provides a fundamental paradigm shift from predictive association to structural causation in financial security, offering actionable insights for the design of resilient, next-generation risk management systems.