Experimental Datasets for Open Source Intelligence Graph Anomaly Detection

Some transactions are labeled as illicit, making the dataset suitable for blockchain anti-money laundering and financial crime detection research.These three datasets differ significantly in terms of node scale, edge scale, feature dimensionality, graph density, and anomaly ratio, thereby covering multiple challenges commonly encountered in open source intelligence graph anomaly detection, such as noisy graph structures, class imbalance, camouflage behaviors, and dynamic evolution.

Specifically, YelpChi contains 45,954 nodes, 3,846,979 edges, 32-dimensional node features, and an anomaly ratio of 14.53%; T-Finance contains 39,357 nodes, 21,222,543 edges, 10-dimensional node features, and an anomaly ratio of 4.58%; Elliptic contains 203,769 nodes, 234,355 edges, 183-dimensional node features, and an anomaly ratio of 9.76%. In the experiments, each dataset is split into training, validation, and test sets at a ratio of 4:3:3, with the random seed fixed to 42 to ensure stability and reproducibility.