Abstract
In recent years, the number of telecom frauds has increased significantly, causing substantial losses to people's daily lives. With technological advancements, telecom fraud methods have also become more sophisticated, making fraudsters harder to detect as they often imitate normal users and exhibit highly similar features. Traditional graph neural network (GNN) methods aggregate the features of neighboring nodes, which makes it difficult to distinguish between fraudsters and normal users when their features are highly similar. To address this issue, we proposed a spatio-temporal graph attention network (GDFGAT) with feature difference-based weight updates. We conducted comprehensive experiments on our method on a real telecom fraud dataset. Our method obtained an accuracy of 93.28%, f1 score of 92.08%, precision rate of 93.51%, recall rate of 90.97%, and AUC value of 94.53%. The results showed that our method (GDFGAT) is better than the classical method, the latest methods and the baseline model in many metrics; each metric improved by nearly 2%. In addition, we also conducted experiments on the imbalanced datasets: Amazon and YelpChi. The results showed that our model GDFGAT performed better than the baseline model in some metrics.