HYBRID GRAPH NEURAL NETWORKS AND XGBOOST FOR FRAUD TRANSACTION DETECTION

Detecting financial fraud in cryptocurrency networks like bitcoin is a significant challenge. The problem is that traditional machine learning models often fail to capture the complex relational patterns within transaction graphs, leading to poor detection. The purpose of this research is to evaluate hybrid models that integrate Graph Neural Networks with XGBoost to improve fraud transaction detection. We conducted a comprehensive analysis on the imbalanced bitcoin dataset, benchmarking standard models (Logistic Regression, Random Forest, Multi-Layer Perceptron) and a standalone XGBoost against two hybrid architectures: Graph Convolutional Network + XGBoost and Graph Attention Network + XGBoost. Our major conclusion is that the hybrid models, particularly Graph Attention Network + XGBoost, achieve a significant improvement. These models effectively leverage both node-level features and the graph's topological structure.