A Noise-Resistant Model for Graph-based Fraud Detection

Graph-based fraud detection is a critical task that identifies anomalous nodes that deviate from the majority of normal nodes within a graph. It can be applied in various practical situations, including but not limited to fake review detection, fraud transaction detection, and bot account detection. Current graph fraud detection models leverage popular Graph Neural Networks (GNNs) as their foundation, achieving significant success from the view of homogeneous and heterogeneous edges. However, these methods assume a sufficient proportion of completely accurate labeled nodes, overlooking the issue of noisy labels present in real-world scenarios. This can lead to significant performance degradation of current graph fraud detection methods. To address this challenge, we propose a Noise-Resistant Model for Graph Fraud Detection. First, we design a foundational graph fraud detection model from a spectral perspective to capture both homogeneous and heterogeneous information of nodes. Based on a conditional variational autoencoder(CVAE), we are able to obtain node features augmented from different perspectives. Next, nodes with noisy labels are trained alongside nodes with clean labels. Utilizing a self-supervised approach, noisy nodes with high prediction confidence that align with their labels are gradually incorporated to the training set. For nodes with lower confidence, we aim to learn better representations and gradually include more of them into the training set. With the augmented features generated by the CVAE, combined with a support set constructed from clean labels, we compute the consistency loss with adversarial strategies to ensure that features augmented from both normal and anomalous perspectives are brought closer to the relevant categories within the support set. Extensive experiments comparing our method with twelve state-of-the-art baselines on six real-world datasets – Amazon, Yelp, Elliptic, FDCompCN, T-Finance, and T-Social – showcase the superiority of our model.