HGsuspector : Scalable Collective Fraud Detection in Heterogeneous Graphs

Graph can straightforwardly represent the relations between the objects, which inevitably draws a lot of attention of both academia and industry. Achievements mainly concentrate on homogeneous graph and bipartite graph. However, it is difficult to use existing algorithm in actual scenarios. Because in the real world, the type of the objects and the relations are diverse and the amount of the data can be very huge. Considering of the characteristics of "black market", we proposeHGsuspector, a novel and scalable algorithm for detecting collective fraud in directed heterogeneous graphs.We first decompose directed heterogeneous graphs into a set of bipartite graphs, then we define a metric on each connected bipartite graph and calculate scores of it, which fuse the structure information and event probability. The threshold for distinguishing between normal and abnormal can be obtained by statistic or other anomaly detection algorithms in scores space. We also provide a technical solution for fraud detection in e-commerce scenario, which has been successfully applied in Jingdong e-commerce platform to detect collective fraud in real time. The experiments on real-world datasets, which has billion nodes and edges, demonstrate that HGsuspector is more accurate and fast than the most practical and state-of-the-art approach by far.