Graph Neural Network Anti-Money-Laundering on the Elliptic Bitcoin Graph

Do graph neural networks find illicit Bitcoin transactions better than classical models? On the Elliptic Data Set (203,769 transactions, 234,355 money flows, 49 time steps) and under the paper's temporal split, a Random Forest on the raw features (test F1 0.797) beats every graph network; the best, GraphSAGE, local, reaches 0.688, and the paired interval of the difference excludes 0. After the dark-market shutdown at step 43 every model falls to F1 near

0. Research demonstration only. It does not accuse any address, transaction or person. The labels are heuristic (they come from the data provider), the features are anonymised so no feature can be named, and no claim is made about real-world performance.

Category: Information Science and Technology

Information Science & Technology Project IST07 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.