Research map / Graph learning / Graph self-supervised learning
Generative Semi-supervised Graph 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 GGAD, a semi-supervised graph anomaly detection method that generates structurally and feature-aligned outlier nodes from labeled normal nodes and trains a one-class classifier with them as negatives.
Problem
Node-level graph anomaly detection when only a small fraction of nodes is labeled as normal (no anomaly labels), where existing unsupervised GAD methods do not fully exploit the normal labels.
Approach
Generate learnable outlier nodes in representation space, initialized from ego networks of labeled normal nodes and optimized with an asymmetric local affinity loss (outliers should have lower local affinity than normal nodes) and an egocentric closeness loss (outliers stay close to their ego-network normal node), then train a binary one-class classifier with them as negatives.
Key results
- GGAD reaches AUROC 0.9443 on Amazon and 0.8228 on T-Finance, outperforming the best semi-supervised baseline across datasets (e.g., AnomalyDAE 0.9171 on Amazon).
- On average over six datasets, GGAD outperforms the best semi-supervised contender AnomalyDAE by about 11% AUROC and 5% AUPRC.
- Ablation shows removing either the local affinity loss or the egocentric closeness loss substantially lowers performance (e.g., Amazon AUROC drops from 0.9324 to 0.8871 without the local affinity loss).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · Amazon | AUROC: 0.9171 -> 0.9443 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · T-Finance | AUROC: 0.6728 -> 0.8228 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · Reddit | AUROC: 0.5829 -> 0.6354 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · Elliptic | AUROC: 0.5409 -> 0.7290 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · Photo | AUROC: 0.6461 -> 0.6476 | No clear effect |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · DGraph | AUROC: 0.5866 -> 0.5943 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · Amazon | AUPRC: 0.7748 -> 0.7922 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · T-Finance | AUPRC: 0.0685 -> 0.1825 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · Reddit | AUPRC: 0.0446 -> 0.0610 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · Elliptic | AUPRC: 0.0949 -> 0.2425 | Helped |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · Photo | AUPRC: 0.1501 -> 0.1442 | Hurt |
| Use GGAD instead of the strongest semi-supervised baseline | Semi-supervised GAD · DGraph | AUPRC: 0.0076 -> 0.0082 | Helped |
Surprising findings
- Random selection of normal nodes as outliers performs poorly, showing that outliers must be distinguishable yet structurally grounded.
- Gaussian perturbation-based outliers give Amazon AUROC of 0.2279, below chance.
Limitations
- The two priors may not capture all anomaly types, so some anomalies may not be represented by generated outliers.
- Performance drops with increasing anomaly contamination in the labeled normal set, though GGAD remains best.
- Declining performance on Amazon and Photo when outlier count grows too large relative to small labeled normal sets.
Experimental setup
Datasets: Amazon T-Finance Reddit Elliptic Photo DGraph
Benchmarks: Amazon T-Finance Reddit Elliptic Photo DGraph
Metrics: AUROC AUPRC
Baselines: DOMINANT AnomalyDAE OCGNN AEGIS GAAN TAM
Builds on: TAM AEGIS DOMINANT AnomalyDAE OCGNN
Closest papers by meaning
- Open-Set Graph Anomaly Detection via Normal Structure Regularisation (ICLR 2025 · similarity 0.91)
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.91)
- ADA-GAD: Anomaly-Denoised Autoencoders for Graph Anomaly Detection (AAAI 2024 · similarity 0.89)
- Semi-supervised Graph Anomaly Detection via Robust Homophily Learning (NeurIPS 2025 · similarity 0.89)
- DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector (ICLR 2025 · similarity 0.88)
- Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark (ICLR 2025 · similarity 0.88)
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.