Fraud detection is a critical challenge in financial transactions, cybersecurity, and e- commerce. Traditional machine learning (ML) techniques often struggle with complex, evolving fraud patterns hidden within large-scale, interconnected data. Graph Neural Networks (GNNs) provide a powerful approach by leveraging graph structures to capture intricate relationships among entities such as users, transactions, and networks. This paper explores how GNNs, combined with advanced ML techniques and optimization strategies, enhance fraud detection capabilities. By modelling transactions as graphs, GNNs enable the identification of suspicious patterns and anomalous behaviours that would be difficult to detect using conventional methods. Furthermore, optimization techniques such as reinforcement learning and hyperparameter tuning improve model efficiency and accuracy. We present a comparative study of different GNN architectures and optimization approaches applied to real-world fraud datasets. The results demonstrate that GNN-based fraud detection significantly outperforms traditional ML models in detecting sophisticated fraudulent activities while minimizing false positives. These findings highlight the potential of GNNs in transforming fraud detection systems across industries.
