An Adaptive Graph Neural Networks Based on Cost-sensitive Learning for Fraud Detection

The world has entered the information age, and intelligence and informatization have become mainstream. The fraudulent means of fraudsters have also become more and more intelligent. Among them, telecommunication fraud has become the main means of fraud. Telecom fraud becomes more and more difficult to prevent with the progress of the times. Therefore, a more innovative and intelligent solution is urgently needed. This paper proposes an adaptive telecom fraud detection model based on a graph neural network (GNN). The model uses an adaptive rate of change to dynamically adjust the step size when finding the optimal sampling threshold. The attention score is weighted with the convolved feature vector. Make the model more attentive to neighboring nodes important to the current task. This model improves its expressiveness and generalization performance. The model is based on call detail records (CDR), modeling users and communication behaviors as nodes and edges of a graph and realizing automatic detection of telecom fraud by learning complex relationships among nodes.