Graph Anomaly Detection via Structure to Attribute Reconstruction

Xingshen Wei, Wei Liu, Wenzhong Li, Sanglu Lu

Abstract: Graph anomaly detection has attracted great attention with wide applications. One mainstream approach for graph anomaly detection is built upon the graph reconstruction framework. However, we observe that existing work mainly relies on same-modal reconstruction (e.g., reconstructing attributes from attributes), which may be less effective in more complex cases. This paper presents a new graph anomaly detection method via cross-modal reconstruction. Unlike existing work, the key idea of this work is to reconstruct node attributes from graph structure. This design enables the detection model to better understand the correlations between node attributes and graph structure, thus being more effective to various types of anomalies. We evaluate the proposed method on four real-world datasets, three types of anomalies, and against 15 existing baselines. The results demonstrate the effectiveness of the proposed method.

External IDs:dblp:conf/icmcs/WeiLLL25