DynGraphTrans: Dynamic Graph Embedding via Modified Universal Transformer Networks for Financial Transaction Data

Abstract

Dynamic graphs representation learning has gradually become a research trend, especially for unsupervised graph embedding learning for numerous graph analytic tasks such as node classification, graph mining and visualization etc. In this paper, we propose a dynamic embedding method, DynGraphTrans, which leverages powerful modelling capability of universal transformer for temporal evolutionary patterns of financial transaction graphs. Real-world transaction graphs are dynamic and continuously evolving over time. According to the characteristics of transaction data, DynGraphTrans computes account node representations by employing a smoothed attention layer and time an interval-aware relative position embedding mechanism. We demonstrate the advantages of the proposed dynamic network embedding by conducting empirical evaluations on a large synthetic transaction graph dataset for an Anti-Money Laundering (AML) task.