Reinforced Cost-Sensitive Graph Network for Detecting Fraud Leaders in Telecom Fraud
Telecommunication fraud has led to significant economic losses worldwide. Although Graph Neural Networks (GNNs) have shown potential in fraud detection, their performance in telecom fraud detection remains suboptimal. The primary reason for this inadequacy is the inability of previous methods to det...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10643627/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Telecommunication fraud has led to significant economic losses worldwide. Although Graph Neural Networks (GNNs) have shown potential in fraud detection, their performance in telecom fraud detection remains suboptimal. The primary reason for this inadequacy is the inability of previous methods to detect “fraud leaders” within fraudulent groups. These leaders orchestrate scams through intermediaries rather than directly engaging in fraudulent activities. Additionally, the significantly smaller number of fraudulent users than benign users further complicates the detection process. To address these challenges, this paper proposes a Reinforced Cost-sensitive Graph Network (RCGN) for detecting fraud leaders in telecom fraud. The model first constructs a graph from users’ call detail records (CDR) and then generates predictions using three base classifiers: GCN, GAT, and GraphSAGE. During the training of the base classifiers, Adaptive Cost-Sensitive Learning (AdaCost) is employed to prioritize fraudulent nodes. Finally, weight coefficients are dynamically optimized using the Deep Deterministic Policy Gradient (DDPG) algorithm, and the prediction results of the three base classifiers are combined to produce the final classification outcomes. Experimental results on extensive real-world datasets demonstrate the effectiveness of RCGN, as it outperforms state-of-the-art GNN and GNN-based fraud detection models. |
|---|---|
| ISSN: | 2169-3536 |