Research map / Graph learning / Graph self-supervised learning
Self-Discriminative Modeling for Anomalous Graph 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.
Trains a GIN-based discriminator on normal graphs alongside a jointly learned pseudo-anomalous graph generator (GAN, VGAE perturbation, or non-adversarial) for unsupervised graph-level anomaly detection.
Problem
Graph-level anomaly detection when only normal graphs are available for training and no real anomalous graphs can be used; existing one-class and distillation methods rely on hypersphere or score assumptions and show performance flips.
Approach
Jointly learn a generator that produces pseudo-anomalous graphs interpolating between normal graphs and real anomalies, and a GIN classifier that separates them from normal graphs; the classifier serves as the anomaly detector. Three variants: adversarial GAN (SDGG-ATI), VGAE perturbation with adversarial training (SDGG-ATII), and non-adversarial joint training (SDGG-NAT).
Key results
- SDGG-NAT reaches 100.00% AUC on MUTAG class 0 and 95.91% on PROTEINS class 0, outperforming the best baseline by over 20% on MUTAG and PROTEINS per the authors.
- On large-scale imbalanced datasets, SDGG-NAT achieves 94.26% (SW-620), 94.20% (MOLT-4), 97.09% (PC-3), 94.71% (MCF-7) AUC, above supervised iGAD (85.82, 83.59, 86.04, 83.22).
- SDGG-NAT has the best AUC in 23 of 27 settings and std at most 5% in 24 of 27 settings in the overall analysis.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SDGG-NAT instead of SDGG-ATI | SDGG graph anomaly detection · PROTEINS (class 0 normal) | AUC: 90.84$\pm$0.15 -> 95.91$\pm$2.55 | Helped |
| Use SDGG-NAT instead of SDGG-ATII | SDGG graph anomaly detection · PROTEINS (class 1 normal) | AUC: 89.19$\pm$0.56 -> 96.26$\pm$0.05 | Helped |
| Replace adversarial training with non-adversarial classifier-generator training | SDGG on large-scale imbalanced graph datasets · SW-620 | AUC: 90.19$\pm$8.94 -> 94.26$\pm$2.86 | Helped |
| Replace SDGG-ATI with SDGG-ATII | SDGG graph anomaly detection · MOLT-4 | AUC: 90.25$\pm$7.57 -> 97.05$\pm$2.39 | Helped |
| Use SDGG-NAT instead of SDGG-ATII | SDGG graph anomaly detection · MOLT-4 | AUC: 97.05$\pm$2.39 -> 94.20$\pm$4.79 | Hurt |
| Train SDGG-NAT rather than SDGG-ATI | SDGG across graph benchmarks · All benchmarks (summary across 27 cases) | Best AUC cases: 3/27 -> 23/27 | Helped |
| Use non-adversarial SDGG-NAT rather than adversarial SDGG-ATI | SDGG across graph benchmarks · All benchmarks (summary across 27 cases) | Cases with std$\leq$5\%: 18/27 -> 24/27 | Helped |
| Use SDGG methods instead of state-of-the-art baselines | Graph-level anomaly detection · MUTAG and PROTEINS | AUC | Helped |
| Use SDGG-ATII instead of supervised iGAD | Large-scale imbalanced graph anomaly detection · MOLT-4 | AUC: 83.59$\pm$1.07 -> 97.05$\pm$2.39 | Helped |
| Use SDGG-NAT instead of supervised iGAD | Large-scale imbalanced graph anomaly detection · MCF-7 | AUC: 83.22$\pm$0.64 -> 94.71$\pm$2.13 | Helped |
Surprising findings
- Fully unsupervised SDGG outperforms supervised graph-level anomaly detection baselines such as DGCNN and iGAD by more than 10% on most large imbalanced datasets.
- Supervised methods trained on limited labeled anomalies may fail to generalize to unseen anomalies on the other side of the decision boundary.
Limitations
- Does not use any real anomalous graphs during training, which may be available in some scenarios.
- SDGG-ATI and SDGG-ATII suffer from min-max training instability; the non-adversarial variant is needed for stability.
- Baselines were reproduced by the authors, and the paper gives a 2D toy analysis rather than a graph-based theory.
Experimental setup
Datasets: MUTAG AIDS COX2 ER_MD PROTEINS DD ENZYMES IMDB-Binary COLLAB SW-620 MOLT-4 PC-3
Benchmarks: MUTAG AIDS COX2 ER_MD PROTEINS DD ENZYMES IMDB-Binary COLLAB SW-620 MOLT-4 PC-3
Metrics: AUC AUCPR F1-score recall
Baselines: SP WL NH RW VGAE-AD OCGIN InfoGraph GLocalKD OCGTL GCN DGCNN GIN
Builds on: GIN VGAE GLocalKD OCGTL iGAD
Closest papers by meaning
- Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection (ICML 2025 · similarity 0.91)
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.89)
- Rayleigh Quotient Graph Neural Networks for Graph-level Anomaly Detection (ICLR 2024 · similarity 0.88)
- ADA-GAD: Anomaly-Denoised Autoencoders for Graph Anomaly Detection (AAAI 2024 · similarity 0.88)
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.87)
- UniGAD: Unifying Multi-level Graph Anomaly Detection (NeurIPS 2024 · similarity 0.87)
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.