Research map / Graph learning / Graph self-supervised learning
Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection
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 AGDiff, which pre-trains a variational graph autoencoder, then uses a conditioned latent diffusion model to generate pseudo-anomalous graphs from normal ones and jointly trains a GIN classifier to separate them from normal graphs for unsupervised graph-level anomaly detection.
Problem
Graph-level anomaly detection suffers from scarce labeled anomalies, since training data usually contains only normal graphs or very few anomalies, which makes learning robust decision boundaries hard.
Approach
Use a latent diffusion model with a learnable noise-perturbation condition to generate pseudo-anomalous graphs close to normal ones in a pre-trained graph latent space, and jointly optimize the diffusion model and a GIN anomaly classifier so the detector learns explicit decision boundaries instead of relying only on reconstruction error.
Key results
- AGDiff reaches 95.83% AUC on MUTAG and 91.21% AUC on ER_MD in Table 1, the best among the compared methods.
- On large imbalanced datasets, AGDiff achieves 91.60% AUC on SW-620 and 94.32% AUC on PC-3, improving over iGAD by 5.78 and 8.28 points.
- Ablation on MUTAG shows AUC falls from 95.83% to 48.16% when the latent diffusion module is removed and reconstruction error is used instead.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pre-training (ablation) | AGDiff · MUTAG | AUC: 71.76±26.65 -> 95.83±2.15 | Helped |
| add pre-training (ablation) | AGDiff · MUTAG | F1-Score: 72.22±27.21 -> 89.45±1.37 | Helped |
| add pre-training (ablation) | AGDiff · ER MD | AUC: 64.10±9.87 -> 91.21±1.84 | Helped |
| add pre-training (ablation) | AGDiff · ER MD | F1-Score: 61.49±9.54 -> 86.04±2.26 | Helped |
| add condition embedding (ablation) | AGDiff · MUTAG | AUC: 63.62±38.32 -> 95.83±2.15 | Helped |
| add condition embedding (ablation) | AGDiff · MUTAG | F1-Score: 65.19±37.44 -> 89.45±1.37 | Helped |
| add condition embedding (ablation) | AGDiff · ER MD | AUC: 68.40±0.21 -> 91.21±1.84 | Helped |
| add condition embedding (ablation) | AGDiff · ER MD | F1-Score: 65.19±0.60 -> 86.04±2.26 | Helped |
| add latent diffusion instead of using reconstruction error (ablation) | AGDiff · MUTAG | AUC: 48.16±4.46 -> 95.83±2.15 | Helped |
| add latent diffusion instead of using reconstruction error (ablation) | AGDiff · MUTAG | F1-Score: 50.67±1.89 -> 89.45±1.37 | Helped |
| add latent diffusion instead of using reconstruction error (ablation) | AGDiff · ER MD | AUC: 63.88±3.07 -> 91.21±1.84 | Helped |
| add latent diffusion instead of using reconstruction error (ablation) | AGDiff · ER MD | F1-Score: 58.48±1.54 -> 86.04±2.26 | Helped |
Surprising findings
- Removing the latent diffusion module and falling back to reconstruction-based scoring drops MUTAG AUC to 48.16%, near chance.
Limitations
- Assumes the normal graph distribution is sufficiently representative, which may not hold in shifting or heterogeneous settings.
- Limited to static graphs; does not handle dynamic graphs.
Experimental setup
Datasets: MUTAG DD COX2 ER_MD SW-620 MOLT-4 PC-3 MCF-7
Benchmarks: MUTAG DD COX2 ER_MD SW-620 MOLT-4 PC-3 MCF-7
Metrics: AUC F1-Score
Baselines: SP kernel WL kernel NH kernel RW kernel OCGIN OCGTL GLocalKD iGAD SIGNET MUSE DO2HSC
Builds on: DDPM GIN GLocalKD iGAD SIGNET
Closest papers by meaning
- DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector (ICLR 2025 · similarity 0.92)
- Self-Discriminative Modeling for Anomalous Graph Detection (ICML 2025 · similarity 0.91)
- 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 0.89)
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.89)
- UniGAD: Unifying Multi-level Graph Anomaly Detection (NeurIPS 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.