<![CDATA[The increasing volume of blockchain transactions has raised significant concerns regarding the detection of irregular and high-risk activities within decentralized financial ecosystems. Conventional anomaly detection approaches tend to focus on transactional values alone, often neglecting the structural relationships that define user interactions. This study introduces a network-based anomaly detection framework that integrates graph embedding and density-based clustering techniques to identify abnormal transaction behaviours. Using a real-world blockchain transaction dataset consisting of 1,316 unique addresses (nodes) and 2,709 transaction links (edges), a directed network model was constructed to represent the flow of digital assets between users. A Singular Value Decomposition (SVD)-based graph embedding was employed to map network structures into a two-dimensional latent space, followed by DBSCAN clustering to isolate low-density outliers. The results indicate that approximately 34 nodes, or 2.6% of the total, were classified as anomalous, exhibiting unusually high transaction volumes, disproportionate connectivity, or bridging characteristics across distinct communities. These findings demonstrate that combining topological representation learning with unsupervised clustering effectively reveals hidden patterns of irregularity within blockchain networks. The proposed framework provides a computationally efficient and interpretable foundation for future integration with advanced graph learning models, such as Graph Neural Networks (GNN), to enhance fraud detection and risk assessment in decentralized systems.]]>
