Research map / Graph learning / Graph self-supervised learning
UniFORM: Towards Unified Framework for Anomaly Detection on Graphs
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 UniFORM, a self-supervised framework that unifies node-, edge-, and graph-level anomaly detection on graphs using dual subgraph sampling, an energy-based GNN, Langevin-dynamics phantom samples, and label-free contrastive meta-learning.
Problem
Graph anomaly detection is usually handled separately at node, edge, and graph levels, and suffers from label sparsity and severe class imbalance, which limits detection of collective anomalies.
Approach
Converts node, edge, and graph inputs into a unified pool of subgraph samples (ego-graphs plus random-walk fragments), scores them with an energy-based GNN, and trains with label-free contrastive meta-tasks where Langevin dynamics generates phantom samples as proxies for anomalies.
Key results
- UniFORM achieves the best node-level AUC on most datasets, with a maximum gain of 4.51% over baselines on BlogCatalog.
- Edge-level AUC improves over BOURNE by up to 2.89%, and graph-level AUC improves by up to 4.27% over the best baseline.
- Ablation on IBM-AML shows removing either the intra-level or inter-level loss significantly hurts performance across all three levels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use UniFORM instead of the strongest reported baseline | Node-level anomaly detection · Cora | AUC: 0.8779 -> 0.9246 | Helped |
| Use UniFORM instead of the strongest reported baseline | Node-level anomaly detection · Pubmed | AUC: 0.9672 -> 0.9428 | Hurt |
| Use UniFORM instead of the strongest reported baseline | Node-level anomaly detection · BlogCatalog | AUC: 0.8184 -> 0.8635 | Helped |
| Use UniFORM instead of the strongest reported baseline | Edge-level anomaly detection · Cora | AUC: 0.9394 -> 0.9487 | Helped |
| Use UniFORM instead of the strongest reported baseline | Edge-level anomaly detection · Enron | AUC: 0.9679 -> 0.8756 | Hurt |
| Use UniFORM instead of the strongest reported baseline | Graph-level anomaly detection · COLLAB | AUC: 0.7208 -> 0.7472 | Helped |
| Remove intra-level loss and replace Langevin dynamics with random feature perturbation (ablation) | UniFORM · IBM-AML node-level | AUC | Hurt |
| Remove intra-level loss and replace Langevin dynamics with random feature perturbation (ablation) | UniFORM · IBM-AML edge-level | AUC | Hurt |
| Remove intra-level loss and replace Langevin dynamics with random feature perturbation (ablation) | UniFORM · IBM-AML graph-level | AUC | Mixed |
| Remove inter-level loss and replace meta-learning with independent encoder training (ablation) | UniFORM · IBM-AML node-level | AUC | Hurt |
| Remove inter-level loss and replace meta-learning with independent encoder training (ablation) | UniFORM · IBM-AML edge-level | AUC | Hurt |
| Remove inter-level loss and replace meta-learning with independent encoder training (ablation) | UniFORM · IBM-AML graph-level | AUC | Hurt |
Surprising findings
- On Pubmed, SL-GAD outperforms UniFORM at node level, which the authors attribute to the high number of injected anomalies.
Limitations
- Node-level gains are not uniform; SL-GAD beats UniFORM on Pubmed, attributed to the high injected anomaly count.
- Graph-level and edge-level results depend on adapting multi-graph datasets (COLLAB, Yelp, IBM-AML) to all levels, and many baselines are not applicable to all datasets.
Experimental setup
Datasets: Cora Pubmed COLLAB BlogCatalog Flickr Enron IMDB Reddit Yelp IBM-AML
Benchmarks: Cora Pubmed COLLAB BlogCatalog Flickr Enron IMDB Reddit Yelp IBM-AML
Metrics: AUC
Baselines: ANOMALOUS DOMINANT CoLA SL-GAD BOURNE AAGAE AAGCN SEAL RGSE InfoGraph-iF GraphCL-iF OCGTL
Builds on: CoLA BOURNE GOOD-D Langevin dynamics GraphCL
Closest papers by meaning
- Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.90)
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.90)
- UniGAD: Unifying Multi-level Graph Anomaly Detection (NeurIPS 2024 · similarity 0.90)
- Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection (ICML 2025 · similarity 0.89)
- Self-Discriminative Modeling for Anomalous Graph Detection (ICML 2025 · similarity 0.89)
- ADA-GAD: Anomaly-Denoised Autoencoders for Graph Anomaly Detection (AAAI 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.