Research map / Graph learning / Graph self-supervised learning

UniGAD: Unifying Multi-level Graph Anomaly Detection

NeurIPS 2024 · Graph learning › Graph self-supervised learning · approach: Graph anomaly detection · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UniGAD, a unified framework for joint node-, edge-, and graph-level graph anomaly detection, using a Rayleigh-quotient-maximizing subgraph sampler (MRQSampler) and a GraphStitch network for multi-level training.

Problem

Graph anomaly detection is usually handled per object level (node, edge, or graph), ignoring correlations among levels, and existing multi-task graph prompt methods do not sample subgraphs suited to anomaly information or balance conflicting multi-level training goals.

Approach

Convert node and edge tasks into graph-level tasks via a DP-based subgraph sampler that greedily maximizes the Rayleigh quotient (spectral anomaly energy) over rooted subtrees, then train identical per-level networks combined by learnable GraphStitch linear mixing of their outputs, with gradient surgery (PCGrad-style projection) to balance per-level losses.

Key results

• UniGAD-BWG reaches 97.84 AUROC on Amazon node-level and 92.18 edge-level, and 86.23 node-level on Yelp, outperforming single-level baselines in most settings.
• In zero-shot transfer, UniGAD-GCN reaches 98.31 AUROC on Weibo node-to-edge, far above GraphPrompt-U and All-in-One-U, which are near 50.
• Ablation on BM-MS shows removing GraphStitch reduces graph AUROC from 99.67 to 95.51 and removing subgraph sampling (w/o GS) reduces graph AUROC to 98.99.

Measured effects reported in the paper

Surprising findings

• Multi-task prompt baselines GraphPrompt-U and All-in-One-U perform near chance (around 50 AUROC) on many node/edge settings.
• Using all 2-hop neighbors as the subgraph sampler can be worse than no subgraph sampling on some datasets, attributed to loss of anomaly information.

Limitations

• The GNN encoder is pre-trained on node-level features only, so edge- and graph-level information may be lost.
• Edge labels for single-graph datasets are synthesized from node anomaly probabilities, and T-Group node labels are derived by a heuristic (appearing in 3 malicious groups).
• UniGAD sometimes sacrifices single-level performance to maintain multi-task performance.

Experimental setup

Datasets: Reddit Weibo Amazon Yelp Tolokers Questions T-Finance BM-MN BM-MS BM-MT MUTAG MNIST0

Benchmarks: Reddit Weibo Amazon Yelp Tolokers Questions T-Finance BM-MN BM-MS BM-MT MUTAG MNIST0

Metrics: AUROC Macro F1 AUPRC

Baselines: GCN GIN GraphSAGE SGC GAT BernNet PNA AMNet BWGNN OCGIN OCGTL GLocalKD

Builds on: GraphMAE BWGNN GraphPrompt All-in-One Stitch Networks RQGNN

Closest papers by meaning

• ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.90)
• UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.90)
• DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector (ICLR 2025 · similarity 0.90)
• Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.90)
• Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark (ICLR 2025 · similarity 0.90)
• Open-Set Graph Anomaly Detection via Normal Structure Regularisation (ICLR 2025 · similarity 0.89)