Inductive Anomaly Detection in Dynamic Graphs with Accumulative Causal Walk Alignment
Résumé
Leveraging structural information and temporal evolution, dynamic graph-based methods have advanced anomaly detection, addressing challenges in real-world social and transaction networks. Despite progress, existing approaches, particularly those based on Graph Neural Networks, face challenges with long-range dependencies and encoding unseen nodes, affecting model performance and generalisation. This paper presents the Dual-Contextual Inductive Dynamic Graph Transformer (DCIDGT), an architecture that efficiently encodes both global and local contexts of target edges. Through a novel mechanism, Accumulative Causal Walk Alignment, DCIDGT captures global spatio-temporal information for unseen nodes and ensures their semantic alignment across snapshots. Our approach is evaluated on real-world cryptocurrency transaction datasets with PR AUC metric that significantly outperforms existing baselines, demonstrating its effectiveness and potential in dynamic graph representation learning.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |
