Bi-Level Selection via Meta Gradient for Graph-Based Fraud Detection

Abstract

Graph Neural Networks (GNNs) have achieved remarkable successes by utilizing rich interactions in network data. When applied to fraud detection tasks, the scarcity and concealment of fraudsters bring two challenges: class imbalance and label noise. In addition to overfitting problem, they will compromise model performance through the message-passing mechanism of GNNs. For a fraudster in a neighborhood dominated by benign users, its learned representation will be distorted in the aggregation process. Noises will propagate through the topology structure as well. In this paper, we propose a Bi-Level Selection (BLS) algorithm to enhance GNNs under imbalanced and noisy scenarios observed from fraud detection. BLS learns to select instance-level and neighborhood-level valuable nodes via meta gradient of the loss on an unbiased clean validation set. By emphasizing BLS-selected nodes in the model training process, bias towards majority class (benign) and label noises will be suppressed. BLS can be applied on most GNNs with slight modifications. Experimental results on two real-world datasets demonstrate that BLS can significantly improve GNNs performance on graph-based fraud detection.