Self-Designing Agentic Workflows for Few-Shot Graph Anomaly Detection
A novel framework named SignGAD (self-designing agentic workflows for few-shot graph anomaly detection) has been introduced to tackle two major issues in graph anomaly detection: the rigidity of fixed pipelines that hinder adaptability across various tasks with limited supervision, and insufficient evidence that obstructs the integration of contextual and structural anomaly signals. SignGAD shifts the focus from training a static anomaly detector to creating detection workflows tailored to specific tasks. It identifies appropriate graph encodings and detector designs to leverage task-relevant anomaly evidence. Additionally, the framework presents a safeguarded final re... (source text truncated). The paper can be found on arXiv with ID 2605.27470.
Key facts
- SignGAD is a novel framework for few-shot graph anomaly detection.
- It addresses fixed pipelines and weak evidence challenges.
- It reformulates anomaly detection as designing task-conditioned workflows.
- It selects graph encodings and detector designs per task.
- It introduces a guarded final re... (truncated).
- The paper is on arXiv: 2605.27470.
Entities
Institutions
- arXiv