Belangrijkste concepten
Guarding graph neural networks from the negative impact of unknown graph anomalies can significantly improve unsupervised graph anomaly detection performance.
Samenvatting
The paper proposes a framework called Guarding Graph Neural Network for Unsupervised Graph Anomaly Detection (G3AD) to address the limitations of existing unsupervised graph anomaly detection methods.
Key highlights:
- Existing GNN-based unsupervised graph anomaly detection methods directly apply GNNs to learn node representations, disregarding the negative impact of graph anomalies on GNNs. This leads to suboptimal node representations and anomaly detection performance.
- G3AD introduces two auxiliary encoders along with correlation constraints to guard the GNNs from encoding inconsistent information induced by graph anomalies.
- G3AD further integrates an adaptive caching module to guard the GNNs from directly reconstructing the observed graph with anomalies, which can provide misleading objectives.
- G3AD comprehensively considers local attribute/topology reconstruction and global consistency alignment for anomaly scoring, enabling effective detection of different types of anomalies.
- Extensive experiments on both synthetic and real-world datasets demonstrate that G3AD outperforms seventeen state-of-the-art unsupervised graph anomaly detection methods.
Statistieken
"Nodes with larger scores are more likely to be considered as anomalies."
"The value of p is fixed as 15 and the value of q is set to 5, 5, 20, 20, 15 for Cora, Citeseer, Pubmed, ACM, and Flickr, respectively."
"The value of k is set to 50."
Citaten
"Most existing methods directly employ GNNs to learn representations, disregarding the negative impact of graph anomalies on GNNs, resulting in sub-optimal node representations and anomaly detection performance."
"To answer the above under-explored research question, in this paper, we propose a simple yet effective framework for Guarding Graph Neural Network for Unsupervised Graph Anomaly Detection (G3AD)."
"Extensive experiments on both synthetic and real-world datasets demonstrate that G3AD outperforms seventeen state-of-the-art unsupervised graph anomaly detection methods."