Answers are grounded in this paper and cite their sources.
Graph neural networks have proved their advanced capabilities on graph-structured data and this technique can be utilized for cryptocurrency graph analysis. However, only a few works concentrate on link prediction throughout the entire bitcoin transaction network (BTN) in real life. All the current work related to Bitcoin link prediction is based on the Bitcoin OTC or Bitcoin Alpha, which are trust-weighted signed networks. The fundamental reason for this is that there are now more than one billion Bitcoin addresses, which makes it challenging for some link prediction algorithms to perform well when applied directly to such a massively sparse network topology. To make training accessible, we shrink the prediction range to a subgraph and adopt GNN methods with stochastic training. We also discovered that the link prediction task cannot fit the BTN where links represent transactions perfectly and proposed a novel method based on the Bitcoin flows to handle the case that the predicted node contains no information. For new address transaction prediction, the RF model achieves the best recall performance for gambling, money laundering, and Ponzi schemes, with recall rates of 94.0%, 96.7%, and 78.1%, respectively. In the case of blackmail prediction, the five models achieve recall rates ranging from 63% to 72%, with the MLP model achieving the highest recall. The best accuracy rates were 92.6% and 97.7% on each task after we trained several models, demonstrating the viability of our suggested approach to resolve the link prediction in BTN where links represent transactions.
No takes yet. Share an insight, caveat, or question.
Yong et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: