Dual perspective-aware graph neural network for graph-level anomaly detection

Abstract

Graph-level anomaly detection based on graph neural networks (GAD-GNN) aims to identify graphs exhibiting anomalous characteristics distinct from the majority in a dataset. However, existing GAD-GNN methods face two critical challenges: Aggregation anomaly dilution occurs when the signals of sparsely distributed abnormal nodes are overwhelmed by the dominant influence of normal nodes during message passing. Readout anomaly dilution arises when locally concentrated anomalies are smoothed out in graph readout. To overcome these challenges, we propose the Dual Perspective-Aware Graph Neural Network (DPGNN), which integrates two complementary modules. The Global Awareness Module enhances node representations with multi-scale return-probability fingerprints, ensuring that signals of sparsely distributed abnormal nodes are preserved against overwhelming normal patterns. The Local Awareness Module adaptively identifies anomaly subgraphs using structural cues and employs attention-based readout to retain concentrated anomalies from being diluted in graph readout. Extensive experiments on multiple benchmark datasets demonstrate that DPGNN consistently outperforms state-of-the-art methods, validating its effectiveness in detecting graph-level anomalies.