A Study on GCN using Focal Loss on Class-Imbalanced Bitcoin Transaction for Anti-Money Laundering Detection

Abstract

Due to a complexity of Bitcoin transaction graph and a class-imbalance difficulty as only a few known labels of blockchain illicit activities exist, graph convolutional network (GCN) using focal loss constraints is applied in this paper for anti-money laundering detection. The GCN, composed of spectral graph convolution layers suitably handles graph structured input data; whereas the focal loss (FL) has been applied to resolve the severe problem of class-imbalance by reducing the loss weight of easy-to-classify samples. In addition, the licit and illicit detection performance comparison on Elliptic dataset among 2, 4 and 6 numbers of GCN layer (GCN -2L, -4L and -6L) relying upon focal loss relative to cross-entropy (CE) loss is assessed. Assessment criteria depend on precision, recall, F1 and accuracy scores, averaged through 10-fold cross validation. A Bitcoin transaction graph, so-called Elliptic dataset derived from Bitcoin blockchain is experimented to detect illicit transactions. The results indicate insignificant difference of accuracy rate between GCN -2L, -4L and -6L models using FL and those employing CE. However, apparently and interestingly, illicit transactions are 4.78%, 5% and 6.20% more correctly predicted by using FL than CE, consecutively relying on GCN -2L, 4L and 6L. Illicit transactions are 4.94%, 5.31% and 7.84% less misclassified by FL than CE.