Research map / Graph learning / Graph self-supervised learning
DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector
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 DiffGAD, a latent-diffusion unsupervised node anomaly detector that distills discriminative content via classifier-free-style guidance between unconditional and common-feature-conditioned diffusion models, with partial-noise reconstruction for detection.
Problem
Unsupervised graph anomaly detection based on reconstruction error often captures shared (common) attributes rather than discriminative content, so anomalies overlap with normal nodes in the latent space.
Approach
Use a graph autoencoder latent space and two diffusion models: an unconditional DM capturing general content and a DM conditioned on a robustly estimated common feature; the discriminative content is the difference between them, injected via a modified score during partial-noise sampling, with reconstruction error as the anomaly score.
Key results
- DiffGAD reaches average ROC-AUC 68.4 ± 1.9 over five datasets, versus 62.7 for the best baseline (ANOMALOUS).
- On Enron, DiffGAD achieves ROC-AUC 71.6 ± 7.0 and reports more than 20.7% AUC gain over GAAN.
- Ablation: unconditional-only diffusion (Diff) reaches average 62.5, close to the AE baseline (61.3), suggesting marginal gain without discriminative distillation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffGAD instead of the strongest listed baseline, Radar | Weibo node anomaly detection · Weibo | ROC-AUC: 98.9 ± 0.1 -> 93.4 ± 0.3 | Hurt |
| Use DiffGAD instead of the strongest listed baseline, LOF | Reddit node anomaly detection · Reddit | ROC-AUC: 57.2 ± 0.0 -> 56.3 ± 0.1 | Hurt |
| Use DiffGAD instead of the strongest listed baseline, IF | Disney node anomaly detection · Disney | ROC-AUC: 57.6 ± 2.9 -> 54.5 ± 0.2 | Hurt |
| Use DiffGAD instead of the strongest listed baseline, AnomalyDAE | Books node anomaly detection · Books | ROC-AUC: 62.2 ± 8.1 -> 66.4 ± 1.8 | Helped |
| Use DiffGAD instead of the strongest listed baseline, GAAN | Enron node anomaly detection · Enron | ROC-AUC: 59.3 ± 0.2 -> 71.6 ± 7.0 | Helped |
| Use DiffGAD instead of the strongest listed baseline, IF | Dgraph node anomaly detection · Dgraph | ROC-AUC: 60.9 ± 0.7 -> 52.4 ± 0.0 | Hurt |
| Add diffusion components to the autoencoder (ablation) | DiffGAD on five-dataset average · Avg (5 datasets) | ROC-AUC: 61.3 ± 2.2 -> 68.4 ± 1.9 | Helped |
| Add discriminative guidance using conditional and unconditional diffusion (ablation) | DiffGAD on five-dataset average · Avg (5 datasets) | ROC-AUC: 62.5 ± 1.0 -> 68.4 ± 1.9 | Helped |
| Replace unconditional diffusion with conditional diffusion alone (ablation) | DiffGAD component comparison · Avg (5 datasets) | ROC-AUC: 62.5 ± 1.0 -> 62.6 ± 0.9 | No clear effect |
| Increase guidance strength to λ = 2.0 (ablation) | DiffGAD · Books, Disney, Enron | ROC-AUC | Helped |
| Set λ = 2 rather than a lower value (ablation) | DiffGAD · Reddit, Weibo | ROC-AUC | Hurt |
| Preserve sample content with timestep t < 500 rather than sample from noise at t = 500 (ablation) | DiffGAD · Books, Disney, Weibo | ROC-AUC | Helped |
Surprising findings
- Plain diffusion on top of the autoencoder gives only marginal improvement over the AE baseline, and conditional and unconditional DMs perform similarly.
- Radar and ANOMALOUS are best on Weibo, outperforming DiffGAD (98.9 vs 93.4 ROC-AUC).
Limitations
- Expressiveness of latent embeddings is limited by the graph encoder.
- Performance is sensitive to the guidance hyperparameter lambda, and the best value differs across datasets.
- Underperforms classical methods on small datasets such as Disney.
Experimental setup
Datasets: Weibo Reddit Disney Books Enron Dgraph
Benchmarks: Weibo Reddit Disney Books Enron Dgraph
Metrics: ROC-AUC AP Recall@k AUPRC
Baselines: LOF MLPAE IF SCAN Radar ANOMALOUS GCNAE DOMINANT DONE AdONE AnomalyDAE GAAN
Builds on: DOMINANT classifier-free guidance EDM sampler denoising score matching
Closest papers by meaning
- Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection (ICML 2025 · similarity 0.92)
- ADA-GAD: Anomaly-Denoised Autoencoders for Graph Anomaly Detection (AAAI 2024 · similarity 0.90)
- Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.90)
- UniGAD: Unifying Multi-level Graph Anomaly Detection (NeurIPS 2024 · similarity 0.90)
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.90)
- Graph Anomaly Detection with Diffusion Model-Based Graph Enhancement (Student Abstract) (AAAI 2024 · similarity 0.90)
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.