SCALE: Adaptive heterogeneous graph prototype learning for open-set unknown financial fraud detection

Financial fraud is becoming more gang-based and evolving dynamically, which exacerbates the ineffectiveness of the existing approaches. Firstly, they struggle to identify gang-based fraud that mimics legitimate transactions. Secondly, they ignore ”pattern drift” caused by evolving fraud tricks. Thirdly, they fail to detect unknown (i.e., previously unseen) fraud patterns in open-set scenarios. To address these challenges, this paper proposes SCALE, an open-set unknown fraud detection framework with adaptive heterogeneous graph prototype learning. Specifically, SCALE first incorporates heterogeneous graphs to model the high-order interconnections among financial entities, and introduces the Graph Structure Entropy (GSE) to examine structural anomalies caused by gang fraud from the information theory perspective. Subsequently, SCALE presents a novel graph representation learning method incorporating the Heterogeneous Graph Variational Information Bottleneck (HG-VIB), which can filter out noise and redundant information to learn informative and task-related fraud features. Finally, SCALE proposes an adaptive prototype learning strategy to automatically update prototypes and capture unknown fraud patterns, which can effectively address fraud pattern drift to detect open-set unknown financial frauds. Experimental results demonstrate that SCALE outperforms state-of-the-art methods, achieving promising robustness and accuracy.