Research: Evaluating the Efficacy of Graph Neural Networks for Fraud Detection in Large...

Research question: How do graph neural network architectures compare to traditional machine learning approaches in detecting fraudulent activity within large-scale transaction networks?

Background & Motivation: Financial fraud remains a persistent threat, with criminals exploiting the complex interconnections in modern transaction networks. Traditional machine learning methods often overlook relational and structural information, which can be critical for accurately identifying fraudulent behavior. Recent advances in graph neural networks (GNNs) offer a promising alternative by leveraging the networked nature of transaction data.

Research Gap / Question: While GNNs have shown potential in various graph-based domains, their effectiveness for fraud detection in real-world transaction networks, compared to established methods, remains underexplored. There is a need for rigorous empirical evaluation and understanding of the conditions under which GNNs outperform conventional classifiers.

Approach & Expected Contribution: This project will review relevant literature, formulate hypotheses, and empirically compare GNN models (e.g., GCN, GraphSAGE) against baseline methods (e.g., random forests, logistic regression) using public transactional datasets, such as IEEE-CIS Fraud Detection. Metrics will include accuracy, precision, recall, and robustness to network sparsity and class imbalance. The project aims to identify key factors influencing GNN performance and provide recommendations for practical deployment.

Significance: By clarifying the strengths and limits of GNNs in this context, this research can guide both academic inquiry and real-world fraud detection strategies, supporting more secure and trustworthy financial systems.