Suspicious: a Resilient Semi-Supervised Framework for Graph Fraud Detection

Graph-based fraud detection is an important task in many real-world domains such as insurance, finance, and cybersecurity. Even if existing semi-supervised models have proven to be efficient in identifying anomalous nodes, they assume that a labeled sample of the nodes is available to train the model, without taking into account the real-world problem of the unreliability of such a sample. In practice, the labeling is often done manually and contains many errors. In this paper, we study fraud detection in attributed networks, and we propose a new framework, based on two graph auto-encoders trained following a suspicion mechanism: the first auto-encoder is trained to better reconstruct the normal nodes while the second one, the fraudulent ones. The final classification is done by coupling the result of both auto-encoders. We demonstrate that our approach obtains at least equivalent performances to state of the art methods in the case of a perfectly labeled sample while being more resilient to the introduction of mistakes in this sample.