An adversarial variational graph autoencoder with contrastive learning for robust anomaly detection in large scale attributed networks

Abstract Anomaly detection in attributed networks is challenging because it combines complex structural relationships with high-dimensional node attributes. Traditional methods rarely solve both problems simultaneously, making them less effective. We present a new approach for anomaly detection that uses adversarial variational graph autoencoders (VGAE) with attention-based residual modeling and contrastive learning to address these issues. The attention-based residual component enhances node embeddings by emphasizing critical connections and preserving essential structural details, improving anomaly detection by retaining subtle yet important information. VGAE is explicitly trained to combine the structure and features into a joint embedding space, while adversarial training prevents overfitting by making embeddings more robust against noise and pushing normal/abnormal nodes further apart. To improve these embeddings, we include the contrastive learning component to move nodes belonging to a community closer and separate them from other communities so that an intrusion or anomaly could be recognized as malicious not only at global levels but inside one or more specific subcommunities at local levels as well. Our method is also capable of dealing effectively with the challenge of imbalanced data in which normal patterns might dominate during training, thus incompetent to detect rare anomalies. Through the execution of extensive experimental investigations utilizing diverse real-world datasets, we establish that our methodology exceeds the performance of the existing premier anomaly detection techniques with respect to both AUC and average precision metrics. These characteristics bestow upon our model a dual advantage of efficacy and scalability, thereby facilitating its applicability to a myriad of practical scenarios, including but not limited to fraud detection in telecommunication networks, spammer hunting in social networks, or identifying errors/omissions of citation information.