Abstract:

Anomaly detection in graph-structured data is essential for applications like social networks and cybersecurity. However, detecting anomalies in static graphs is challenging due to complex relationships, data sparsity, and the lack of labeled data. Traditional methods depend on handcrafted features and often miss complex dependencies, while recent Graph Neural Network (GNN) methods have demonstrated strong capabilities to treat structural and attribute information separately.
To mitigate these constraints, this paper introduces a dual-encoder framework that concurrently models structural and attribute information through a GraphSAGE-based encoder featuring node-level attention and an MLP-based attribute encoder. A hybrid anomaly scoring system that combines structural–attribute inconsistency with neighbor-aware deviation is presented.
Tests on benchmark datasets like Cora, Citeseer, and PubMed show that the proposed method achieves competitive performance, highlighting its effectiveness for anomaly detection in static graphs.