Graph Fraud Detection Based on Accessibility Score Distributions

Graph fraud detection approaches traditionally present frauds as subgraphs and focus on characteristics of the fraudulent subgraphs: unexpectedly high densities or sparse connections with the rest of the graph. However, frauds can easily circumvent such approaches by manipulating their subgraph density or making connections to honest user groups. We focus on a trait that is hard for fraudsters to manipulate: the unidirectionality of communication between honest users and fraudsters. We define an accessibility score to quantify the unidirectionality, then prove the unidirectionality induces skewed accessibility score distributions for fraudsters. We propose SkewA, a novel fraud detection method that measures the skewness in accessibility score distributions and uses it as an honesty metric. SkewA is (a) robust to frauds with low density and various types of camouflages, (b) theoretically sound: we analyze how the unidirectionality brings skewed accessibility score distributions, and (c) effective: showing up to 95.6% accuracy in real-world data where all competitors fail to detect any fraud.