Blockchain technology has introduced a decentralized and transparent mechanism for recording transactions; however, the increasing volume and interconnectivity of blockchain networks also raise the risk of fraudulent and high-risk activities. This study proposes a Graph Neural Network (GNN)-based framework to evaluate the risk levels of blockchain nodes by integrating both transactional attributes and structural relationships. Using a dataset of 10,000 blockchain records and approximately 412,000 edges, the network was modelled as a graph in which each node represents an address and edges denote transaction or similarity links. As baselines, Random Forest and XGBoost models were employed, achieving accuracies of 0.94 and 0.95, respectively, with F1-scores of 0.93 and 0.94. These models effectively captured individual node patterns but lacked awareness of inter-node dependencies. The proposed GNN model demonstrated the highest overall performance, with an accuracy of 0.96 and an F1-score of 0.95, by learning from both node attributes and their topological context. This approach enabled the identification of high-risk nodes that traditional models failed to detect. The results confirm that network-based learning significantly enhances the accuracy and interpretability of blockchain risk analysis. The proposed GNN framework provides a scalable foundation for real-time blockchain monitoring, anomaly detection, and governance systems, contributing to improved transparency and resilience within decentralized financial ecosystems.