AttyAbson/GNN-vs.-XGBoost-for-Fraud-Detection-and-CBN-AML-KYC-Compliance-in-Nigeria-s-Fintech-Sector-

**Comparative Analysis of GNN and XGBoost for Blockchain Fraud Detection (Nigeria Fintech / CBN AML–KYC)** This project investigates how Graph Neural Networks (GNN) and XGBoost can be used to improve fraud detection on blockchain transactions, with a focus on Nigeria’s fintech and CBN AML/KYC compliance. **Project Overview** Domain: Blockchain fraud detection / AML monitoring Context: Nigerian fintech ecosystem, synthetic but regulation-inspired data Core question: Which approach is more effective and practical for transaction monitoring – GNN or XGBoost, enhanced with anomaly detection and RL? **Data & Setup** ~30,000 synthetic blockchain-style transactions (legit vs suspicious). Features reflect real AML patterns: burstiness, mixer usage, stablecoins, CBDC, KYC tiers, etc. Transactions modelled as a graph: accounts = nodes, transfers = edges. **Methods** XGBoost on engineered tabular features. GNN (GCN) on the transaction graph for relational patterns. Anomaly detection (e.g. Isolation Forest) adds anomaly scores. Reinforcement Learning (DQN) tunes decision thresholds / flagging policy. **Key Findings** Base GNN: very strong fraud detection (high recall & precision), best at multi-hop laundering patterns. Base XGBoost: conservative (few false positives) but misses many fraud cases. XGBoost + anomaly detection + RL: reaches near-perfect performance with lower compute cost and better explainability (SHAP). GNN remains powerful for complex network analysis, but heavier to deploy at scale. **Practical Insight** Recommended hybrid architecture: XGBoost + RL for scalable first-line screening. GNN for deep investigation of high-risk clusters and complex fraud rings.