HoME: Homogeneity-Mining-Based Embedding Towards Detecting Illicit Transactions on Bitcoin

Cryptocurrencies have brought booming economic innovations in recent years, but they have also given rise to many intractable problems related to financial crimes, urgently requiring for some effective countermeasures. Transaction network analysis based on machine learning methods has been widely used to detect illicit transactions. In this paper, we aim at mining the homogeneity of transaction patterns, with the inspiration that homogeneous transactions share similar dominant features and such information can be used to enhance illicit detection. Accordingly, we propose the Homogeneity-Mining-based Embedding (HoME) framework. The main idea of HoME is to convert homogeneity among transactions into topological connectivity for representation learning. First, we design the Pure Cluster Search algorithm based on agglomerative hierarchical clustering to find out groups of homogeneous transactions. Then, we represent homogeneity using virtual vertices and add them into the transaction network. Further, we design a Two-Phase Walk method to capture neighborhood similarity in the new transaction network. As shown in evaluations on the well-known Elliptic dataset, HoME can achieve good effectiveness and robustness, outperforming popular baselines.