Telecom Fraud Detection via Imbalanced Graph Learning

In recent years, an increasing number of telecom frauds have caused huge losses to people around the world. Graph neural network(GNN) brings new possibilities for telecom fraud detection. However, due to the existence of the imbalance problem in the graph, it is difficult for general GNN models to detect a small number of positive samples. To address this problem, we design a new GNN-based fraud detector. First, we transform the node features with a multilayer perceptron. Subsequently, a reinforcement learning-based neighbor sampling strategy is designed to balance different classes of node neighbors. Then the node features are aggregated using GNN, while the focal loss function is used to measure the model error. We conducted experiments on two real-world telecom fraud datasets, and the results show that the proposed method does have competitive performance, especially for the graph imbalance problem.