Exploring Graph-Based Machine Learning Techniques for Transaction Fraud Detection: A Comparative Analysis of Performance

No Thumbnail Available
Zade, Nikita Prakash
Issue Date
MSc in Data Analytics
Dublin Business School
This report investigates graph-based fraud detection methods, focusing on credit card transactions to tackle the increasing complexity of financial fraud. It explores various machine learning models viz. Isolation Forest, Random Forest, Graph Autoencoders (GAE), and Graph Convolutional Networks (GCN), assessing them based on evaluation metrics. The Random Forest emerges as a robust model with consistent high performance, while the Isolation Forest shows minimal effectiveness. The GAE and GCN demonstrate potential, especially with hyperparameter tuning. Significant improvements in accuracies were observed post-tuning, particularly with the GCN model, showcasing the importance of model optimization. The research acknowledges the challenges of acquiring graph structured data, real-time analysis, adaptive fraudsters, and data privacy in implementing graph-based fraud detection. Conclusively, the study endorses graph-based methods as a formidable approach to enhance fraud detection, emphasizing continuous research and development to address existing challenges and improve system scalability, efficiency, and security. Accuracies obtained posttuning are notably high for Random Forest and GCN, indicating their effectiveness in fraud detection scenarios.