TAAD: Time-varying adversarial anomaly detection in dynamic graphs

The timely detection of anomalous nodes that can cause significant harm is essential in real-world networks. One challenge for anomaly detection in dynamic graphs is the identification of abnormal nodes at newly emerged moments. Unfortunately, existing methods tend to learn nontransferable features from historical moments that do not generalize well to newly emerged moments. In response to this challenge, we propose Time-varying Adversarial Anomaly Detection (TAAD), a generalizable model to learn transferable features from historical moments, which can transfer prior anomaly knowledge to newly emerged moments. It comprises four components: the feature extractor, the anomaly detector, the time-varying discriminator and the score generator. The time-varying discriminator cooperates with the feature extractor to conduct adversarial training, which decreases the distributional differences in the feature representations of nodes between historical and newly emerged moments to learn transferable features. The score generator measures the distributional differences of feature representations between normal and abnormal nodes, and further learns discriminable features. Extensive experiments conducted with four different datasets present that the proposed TAAD outperforms state-of-the-art methods.