NCTAF-GCN: Node-Contrastive Topology-Adaptive Fairing Graph Convolutional Network for anomaly detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Graph anomaly detection plays an important role in complex systems, like financial system s , social media and computer networks , for identifying malicious components. Many researchers choose graph convolutional networks (GCNs) as the backbone model to capture structural and attribute information. Nevertheless, GCNs suffer from three inherent challenges that lead to inferior performance. Firstly, over-smoothing presentation undermines the downstream task. It means node’s features mix with many neighbors’ features. Secondly, the difference of aggregation capability for nodes. High-degree nodes are prone to capture enough information in shallow layers while low-degree nodes are still lacking common features. Thirdly, GCN is sensitive to noise. Both structure and attribute deviation could cause misclassification. To address these challenges, we propose a Node-Contrastive Topology-Adaptive Fairing Graph Convolutional Network (NCTAF-GCN), a contrastive learning method, for anomaly detection. Specifically, two topology-adaptive graphs are constructed in a heuristic way . The adaptive graphs improve the original graph representation with augment ed information. Additionally, there is a feature-level attention mechanism for enhancing anomalous features. Experiments conducted on five real-world datasets indicate that the detection capability of NCTAF-GCN surpasses that of comparative baselines.