GNNs Refine Bitcoin Address Clustering, Enhance User Activity Analysis.

Analyzing user activity on Bitcoin is challenging due to its pseudonymous nature, where a single user can control multiple addresses. Existing heuristic-based clustering methods attempt to group addresses belonging to the same user but often suffer from limited modularity and errors, such as incorrectly merging different users. This research introduces a novel approach to refine these heuristic-obtained clusters. It grounds the clustering process on contrastive embeddings generated by graph neural networks (GNNs), providing a more nuanced understanding of address relationships. The contributions include releasing a public dataset of Bitcoin transaction graphs with numerous clusters, proposing a methodology for learning address embeddings consistent with heuristics, and enabling hierarchical clustering for finer analysis. This method also provides a quantitative criterion to identify and flag suspicious merges, improving the accuracy of user-level activity analysis.