IDGL: An Imbalanced Disassortative Graph Learning Framework for Fraud Detection

Abstract

AbstractThe thriving growth of Internet service not only facilitates our daily lives but also incubates various fraudulent activities with concealment. The traceable interactive behaviors forming the graph-like data provide a great opportunity for graph-based fraud detection. Owing to the stellar performance of assortative graph learning, GNN-based fraud detection methods escalate much attention. However, the fraud graph is not always assortative but more likely disassortative as the fraudsters usually camouflage themselves via building numerous connections with normal users. Additionally, the GNN-based fraud detection methods also suffer from graph imbalance issues as the number of fraudsters is far less than that of the normal users. To address these problems, an imbalanced disassortative graph learning framework (IDGL) is proposed with two key components. First, an adaptive dual-channel convolution filter is developed to adaptively combine the advantage of low- and high-frequency signals from its neighbors so as to assimilate the nodes with assortative edges and discriminate the nodes with disassortative edges. Second, a label-aware nodes and edges sampler is designed with the consideration of nodes’ popularity and corresponding label class frequency, which helps the model simultaneously eliminate the bias towards the major classes and pay more attention to the valuable connections (fraud-fraud, fraud-benign). Extensive experiments on two public fraud datasets demonstrate the effectiveness of our method. KeywordsFraud detectionGraph disassortativityGraph imbalanceAdaptive frequency filterLabel-aware sampler