Research map / Graph learning / Graph self-supervised learning
ADA-GAD: Anomaly-Denoised Autoencoders for 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 ADA-GAD, a two-stage graph anomaly detection framework that pretrains graph autoencoders on learning-free anomaly-denoised augmented graphs (node, edge, subgraph masking) and then retrains decoders on the original graph with a node anomaly distribution regularizer.
Problem
Unsupervised graph anomaly detection with reconstruction-based GNN autoencoders suffers from Anomaly Overfitting and Homophily Trap, since anomalies in the training graph corrupt reconstruction errors of normal nodes and anomalies themselves get reconstructed well.
Approach
Reduce the anomaly level of the training graph via learning-free masking augmentations selected by a spectral high-frequency energy metric, pretrain multi-level (node, edge, subgraph) graph autoencoders on these cleaner graphs, freeze encoders, retrain decoders on the original graph, and regularize the per-node anomaly score distribution to be sharper.
Key results
- ADA-GAD achieves AUC 84.73 on Cora and 83.25 on Amazon, the best among compared methods on all seven datasets in the reported table.
- On Weibo ADA-GAD reaches 98.44 AUC versus 98.27 for the best baseline Radar/ANOMALOUS, and on Disney it reaches 70.04 versus 68.19 for OCGNN.
- Single-level anomaly-denoised variants (node, edge, subgraph) are slightly below the full three-level model but outperform random augmentation (ADA-GAD_rand), e.g. Books 62.77/62.71/62.70 vs 61.75.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply ADA-GAD instead of the strongest listed baseline | Cora · Cora | AUC (%): 84.34±0.03 -> 84.73±0.01 | Helped |
| apply ADA-GAD instead of the strongest listed baseline | Amazon · Amazon | AUC (%): 82.62±0.26 -> 83.25±0.03 | Helped |
| apply ADA-GAD instead of the strongest listed baseline | Weibo · Weibo | AUC (%): 98.27±0.00 -> 98.44±0.33 | Helped |
| apply ADA-GAD instead of the strongest listed baseline | Reddit · Reddit | AUC (%): 56.64±0.00 -> 56.89±0.01 | Helped |
| apply ADA-GAD instead of the strongest listed baseline | Disney · Disney | AUC (%): 68.19±1.45 -> 70.04±3.08 | Helped |
| apply ADA-GAD instead of the strongest listed baseline | Books · Books | AUC (%): 59.86±4.82 -> 65.24±3.17 | Helped |
| apply ADA-GAD instead of the strongest listed baseline | Enron · Enron | AUC (%): 66.86±0.54 -> 72.89±0.86 | Helped |
| use the combined node-, edge-, and subgraph-level branches rather than random augmentation (ablation) | ADA-GAD · Cora | AUC (%): 81.61±0.01 -> 84.73±0.01 | Helped |
| use the combined node-, edge-, and subgraph-level branches rather than random augmentation (ablation) | ADA-GAD · Amazon | AUC (%): 76.36±0.12 -> 83.25±0.03 | Helped |
| use the combined node-, edge-, and subgraph-level branches rather than random augmentation (ablation) | ADA-GAD · Weibo | AUC (%): 90.74±0.65 -> 98.44±0.33 | Helped |
| use the combined node-, edge-, and subgraph-level branches rather than random augmentation (ablation) | ADA-GAD · Reddit | AUC (%): 56.03±0.38 -> 56.89±0.01 | Helped |
| use the combined node-, edge-, and subgraph-level branches rather than random augmentation (ablation) | ADA-GAD · Disney | AUC (%): 68.56±2.94 -> 70.04±3.08 | Helped |
Surprising findings
- Non-deep methods Radar and ANOMALOUS outperform many deep competitors on Weibo and Reddit, which the authors attribute to overfitting of deep models.
- Training DOMINANT on a cleaner graph (fewer injected anomalies) consistently yields better detection, motivating the anomaly-denoising idea.
Limitations
- Augmentation is learning-free and uses a simplified masking heuristic rather than directly solving the constrained anomaly-minimization problem.
- Hyperparameters such as encoder depth, regularization weight, and augmentation budgets vary per dataset and are tuned on the reported results.
- Evaluated only on node-level anomaly detection with injected or organic anomalies; supervised and graph-level settings are left to future work.
Experimental setup
Datasets: Cora Amazon Weibo Reddit Disney Books Enron
Benchmarks: Cora Amazon Weibo Reddit Disney Books Enron
Metrics: AUC
Baselines: SCAN Radar ANOMALOUS GCNAE DOMINANT DONE AdONE AnomalyDAE GAAN CoLA OCGNN CONAD
Builds on: DOMINANT PyGOD
Closest papers by meaning
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.92)
- DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector (ICLR 2025 · similarity 0.90)
- Rethinking Reconstruction-based Graph-Level Anomaly Detection: Limitations and a Simple Remedy (NeurIPS 2024 · similarity 0.90)
- Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.89)
- Generative Semi-supervised Graph Anomaly Detection (NeurIPS 2024 · 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.