Leveraging Graph Neural Networks to Detect Anomalies in Mobile Money Transactions

Anomaly detection is a critical task across various domains, including data quality management, healthcare, and finance. Several anomaly detection methods particularly rule-based approaches have been designed mainly working with simple data structures. However, given the evolution in data, traditional rule-based techniques often fall short when dealing with complex datasets with intricate interdependencies such as those on networks. These interdependencies provide extra information that could improve the anomaly detection task. This paper, therefore, introduces an enhanced anomaly detection method leveraging Graph Neural Networks (GNNs), which are adept at capturing patterns in data with rich relational structures. Our approach demonstrates superior performance over conventional rule-based models, achieving a balanced accuracy of 93% and an F1 score of 92.3%, compared to 72% and 58.5% respectively for the traditional methods. These results underscore the efficacy of adopting GNNs in accurately identifying anomalies, with significant implications for applications. Hence the use of GNN instead of rule-based methods in anomaly detection problems would yield more accurate and stable results particularly in problems with complex data.