Spade+: A Generic Real-Time Fraud Detection Framework on Dynamic Graphs

Jiaxin Jiang, Yuhang Chen, Bingsheng He, Min Chen, Jia Chen

Abstract: Real-time fraud detection remains a pressing issue for many financial and e-commerce platforms. $\mathsf {Grab}$, a prominent technology company in Southeast Asia, addresses this by constructing a transactional graph. This graph aids in pinpointing dense subgraphs, possibly indicative of fraudster networks. Notably, prevalent methods are designed for static graphs, neglecting the evolving nature of transaction graphs. This static approach is ill-suited to the real-time necessities of modern industries. In our earlier work, $\mathsf {Spade}$, the focus was mainly on edge insertions. However, $\mathsf {Grab}$'s operational demands necessitated managing outdated transactions. Persistently adding edges without a deletion mechanism might inadvertently lead to densely connected legitimate communities. To resolve this, we present $\mathsf {Spade+}$, a refined real-time fraud detection system at $\mathsf {Grab}$. Contrary to $\mathsf {Spade}$, $\mathsf {Spade+}$ manages both edge additions and removals. Leveraging an incremental approach, $\mathsf {Spade+}$ promptly identifies suspicious communities in large graphs. Moreover, $\mathsf {Spade+}$ efficiently handles batch updates and employs edge packing to diminish latency. A standout feature of $\mathsf {Spade+}$ is its user-friendly APIs, allowing for tailored fraud detection methods. Developers can easily integrate their specific metrics, which $\mathsf {Spade+}$ autonomously refines. Rigorous evaluations validate the prowess of $\mathsf {Spade+}$; fraud detection mechanisms powered by $\mathsf {Spade+}$ were up to a million times faster than their static counterparts.