Reproduce “GFM4GA: Graph Foundation Model for Group Anomaly Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jiujiu Chen, Weijun Zeng, Shaofeng Hu, Sihong Xie, Hui Xiong

Researchers from HKUST and Tencent introduce GFM4GA, the first Graph Foundation Model specifically for group anomaly detection. The model leverages dual-level contrastive pretraining and an adaptive fine-tuning strategy, achieving average performance improvements of 2.85% in AUROC and 2.55% in AUPRC over competing baselines in 10-shot learning scenarios.

Key insights

  • Subgraph-level contrastive learning is paramount for learning the fundamental patterns distinguishing an anomalous group core from its surrounding subgraph.
  • Dynamic group anomaly contexts, built from local neighborhood information, are crucial for resolving ambiguities and overcoming the camouflage exhibited by anomalous groups.
  • A feature-deviation-based Graph Neural Network encoder combined with parameter-constrained fine-tuning allows for effective transfer of learned general knowledge while adapting to specific group anomaly types.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.