TY - GEN
T1 - FCMH
T2 - 20th International Conference on Advanced Data Mining Applications, ADMA 2024
AU - Zhang, Rui
AU - Li, Wenbo
AU - Ning, Xiaodong
AU - Cheng, Dawei
AU - Han, Li
AU - Yang, Heguo
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.
PY - 2025
Y1 - 2025
N2 - The issue of financial fraud is attracting increasing social attention. The task of Graph Fraud Detection (GFD) is a typical application of Graph Neural Networks (GNNs) in the e-commerce field. Although several successful works have already achieve excellent performance, there are still many obstacles before these methods can be used in financial market industry. Many of the GFD methods are based on spatial approaches, such as neighbor selection, which will slow down their execution on large datasets. This poses problems in e-commerce settings where real-time detection is essential. Also, fraudsters often disguise their behavior, resulting in high heterophily of the graph, which makes direct aggregation of neighbors less effective. To address these issues, we introduce the Fast Cluster Multi-Hop Detector (FCMH). This model first partitions the graph into different small subgraphs to accelerate the neighbor aggregation process. Then, it uses a multi-hop neighborhood aggregation strategy to simultaneously aggregate multiple layers of neighbors to learn fraudulent patterns. Our model has been evaluated on both the open source and industrial e-commerce datasets, and it has produced ideal results in terms of classification performance and time efficiency. We believe that our work will be beneficial for the applications of GFD models in financial fraud detection.
AB - The issue of financial fraud is attracting increasing social attention. The task of Graph Fraud Detection (GFD) is a typical application of Graph Neural Networks (GNNs) in the e-commerce field. Although several successful works have already achieve excellent performance, there are still many obstacles before these methods can be used in financial market industry. Many of the GFD methods are based on spatial approaches, such as neighbor selection, which will slow down their execution on large datasets. This poses problems in e-commerce settings where real-time detection is essential. Also, fraudsters often disguise their behavior, resulting in high heterophily of the graph, which makes direct aggregation of neighbors less effective. To address these issues, we introduce the Fast Cluster Multi-Hop Detector (FCMH). This model first partitions the graph into different small subgraphs to accelerate the neighbor aggregation process. Then, it uses a multi-hop neighborhood aggregation strategy to simultaneously aggregate multiple layers of neighbors to learn fraudulent patterns. Our model has been evaluated on both the open source and industrial e-commerce datasets, and it has produced ideal results in terms of classification performance and time efficiency. We believe that our work will be beneficial for the applications of GFD models in financial fraud detection.
KW - Deep Learning
KW - Graph Fraud Detection
KW - Graph Neural Networks
UR - https://www.scopus.com/pages/publications/85213379550
U2 - 10.1007/978-981-96-0821-8_3
DO - 10.1007/978-981-96-0821-8_3
M3 - 会议稿件
AN - SCOPUS:85213379550
SN - 9789819608201
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 34
EP - 49
BT - Advanced Data Mining and Applications - 20th International Conference, ADMA 2024, Proceedings
A2 - Sheng, Quan Z.
A2 - Zhang, Xuyun
A2 - Wu, Jia
A2 - Ma, Congbo
A2 - Dobbie, Gill
A2 - Jiang, Jing
A2 - Zhang, Wei Emma
A2 - Manolopoulos, Yannis
A2 - Mansoor, Wathiq
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 3 December 2024 through 5 December 2024
ER -