Mining Mobile Network Fraudsters with Augmented Graph Neural Networks

Area of Science:

• Telecommunications Engineering
• Data Science
• Machine Learning

Background:

• Mobile communication networks generate massive call detail records (CDR), creating opportunities for fraudsters.
• Detecting telecom fraud is challenging due to data volume and inherent graph imbalances.
• Existing graph neural network (GNN) methods struggle with graph imbalance and oversmoothing, limiting fraud detection efficacy.

Purpose of the Study:

• To propose a novel fraud detection framework for mobile communication networks.
• To address the limitations of graph imbalance and GNN oversmoothing in telecom fraud detection.
• To enhance the accuracy and effectiveness of identifying fraudsters within large-scale CDR datasets.

Main Methods:

• User features are transformed using a multilayer perceptron.
• A reinforcement learning-based neighbor sampling strategy is employed for graph balancing.
• Graph neural networks (GNNs) are utilized for feature aggregation.
• An AdaBoost algorithm integrates multiple GNN-based weak classifiers.
• A balanced focal loss function is implemented for model training.

Main Results:

• The proposed method significantly improves fraud detection performance on real-world telecom datasets.
• The approach effectively mitigates the challenges posed by graph imbalance.
• The GNN oversmoothing problem is successfully addressed, leading to more accurate fraud identification.
• Experimental results demonstrate the superiority of the proposed method over existing techniques.

Conclusions:

• The developed fraud detection framework offers a robust solution for identifying fraudsters in mobile networks.
• The combination of GNNs, reinforcement learning, and AdaBoost provides a powerful tool for tackling telecom fraud.
• This research contributes a significant advancement in the field of fraud detection within telecommunication systems.