LabelGen: An Anomaly Label Generative Framework for Enhanced Graph Anomaly Detection

Anomaly detection in graphs is increasingly used to reveal fraud, fakes, security attacks and unusual behaviours in networks, such as social networks, financial transaction networks and the Internet of Things. Accurately detecting such graph anomalies using deep learning approaches faces challenges in terms of obtaining sufficient labelled data, as well as an imbalance between normal and anomalous instances. These contribute to model bias or over-fitting problems and inferior anomaly detection outcomes. In order to address these challenges in graphs, we propose a novel generative framework, called LabelGen, that can generate additional anomalous labels, in terms of graph objects, such as nodes, through augmentation and provide updated deep embedding for the graph concurrently. In particular, we propose the use of a k-hop neighborhood sampling strategy, an anomaly scoring mechanism and an adversarial learning framework with a generator and discriminator pair in order to generate sufficient and informative anomalous nodes that closely resemble the characteristics of existing anomalies in the graph. Evaluation on benchmark network datasets, as well as ablation and comparison studies with random label generation processes and other existing works reveal that the proposed generative framework is superior in improving the anomaly detection accuracy in graphs, while achieving a balanced trade-off between accuracy and computational efficiency.