GUFAD: A Graph-based Unsupervised Fraud Account Detection Framework

Wei Zhang, Yufei Zhang, Yilin Huang, Fangshu Chen, Jiahui Wang, Xiaoming Hu

Abstract: The detection of fraud accounts at scenarios of registration and login is a critical task for Internet enterprises, which can help to avoid economic losses at a very early stage. In industry, most companies tend to deploy supervised models such as rule-based models. However, these methods are significantly restricted in scalability since they highly rely on domain knowledge and manual annotations. Therefore, we designed a novel account graph analysis approach for uncovering fraudulent patterns.The framework explores an advanced feature-account bigraph to calculate the aggregation of accounts and applies a community detection algorithm to detect organized fraud groups. Next, a graph embedding and clustering algorithm is introduced to further analyze the types of aggregated communities, which can help to reduce account misclassification. Furthermore, a novelty method POMV is designed to explore the patterns of missing values. And two dynamic feature aggregation methods based on multi-granularity sliding windows are proposed to construct expressive features that can help to improve the quality of the account graph. The proposed framework achieves 0.82 F1_Score on average, which significantly outperforms the off-the-shelf models by 3% to 16%.

External IDs:dblp:conf/icmlca/ZhangZHCWH23