Research map / Graph learning / Graph self-supervised learning
Revisiting Graph Contrastive Learning on Anomaly Detection: A Structural Imbalance Perspective
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 AD-GCL, a graph contrastive learning framework for unsupervised anomaly detection that improves tail (low-degree) node detection via neighbor pruning on head nodes, anomaly-guided neighbor completion for tail nodes, and intra/inter-view consistency losses.
Problem
Graph contrastive learning anomaly detectors perform well overall but fail to detect tail anomalies (low-degree nodes) under the power-law structural imbalance common in real-world graphs.
Approach
Forge tail nodes from head nodes by similarity-aware neighbor pruning and align head-node knowledge to them across views, while enlarging tail neighborhoods via mixup with auxiliary nodes sampled by anomaly-score and feature similarity, all within a GCL discriminator trained with intra- and inter-view losses.
Key results
- AD-GCL reports AUC of 92.83 on Cora, 94.88 on Citeseer, and 95.74 on Pubmed, best among compared methods on those datasets.
- On Bitcoinotc, BITotc and BITalpha AD-GCL reaches AUC 82.19, 82.11 and 79.62, outperforming runner-up methods.
- Ablation shows removing intra-view contrast loss causes the largest tail-node drop (30.62%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AD-GCL instead of the strongest baseline | Graph anomaly detection · Cora, Citeseer, Pubmed, Bitcoinotc, BITotc, BITalpha | AUC: outperforms the runner-up methods by 2.13%, 2.84%, 0.45%, 2.89%, 1.21%, and 2.71%, respectively | Helped |
| add neighbor pruning and head-to-forged-tail alignment (ablation) | AD-GCL · Tail nodes across six datasets | AUC: 2.46% | Helped |
| add anomaly-guided neighbor completion (ablation) | AD-GCL · Tail nodes across six datasets | AUC: 1.87% | Helped |
| add intra-view contrast loss (ablation) | AD-GCL · Tail nodes across six datasets | AUC: 30.62% | Helped |
| add inter-view contrast loss (ablation) | AD-GCL · Tail nodes across six datasets | AUC: 1.89% | Helped |
| replace multinomial sampling with k-nearest neighbor sampling (ablation) | AD-GCL · Cora, Citeseer, Pubmed, Bitcoinotc, BITotc, BITalpha | AUC: 92.83 vs 90.33; 94.88 vs 93.15; 95.74 vs 95.36; 82.19 vs 80.38; 82.11 vs 80.85; 79.62 vs 79.08 | Hurt |
| replace neighbor pruning and completion with random edge dropping (ablation) | AD-GCL · Cora, Citeseer, Pubmed | AUC: 92.83 vs 91.91; 94.88 vs 91.82; 95.74 vs 95.01 | Hurt |
| replace neighbor pruning and completion with feature masking (ablation) | AD-GCL · Cora, Citeseer, Pubmed | AUC: 92.83 vs 91.63; 94.88 vs 92.95; 95.74 vs 94.64 | Hurt |
| replace anomaly-guided completion with naive edge completion (ablation) | AD-GCL · Cora, Citeseer, Pubmed | AUC: 92.83 vs 78.28; 94.88 vs 70.33; 95.74 vs 92.03 | Hurt |
| replace MEAN readout with MAX or MIN (ablation) | AD-GCL · Cora, Citeseer, Pubmed, Bitcoinotc, BITotc, BITalpha | AUC: MAX: 91.54, 94.08, 94.71, 81.81, 81.28, 78.11; MIN: 92.57, 93.77, 95.36, 81.28, 79.76, 78.34; MEAN: 92.83, 94.88, 95.74, | Mixed |
| increase sampling rounds R (ablation) | AD-GCL · Cora | AUC | Helped |
| increase hidden dimension beyond 64 (ablation) | AD-GCL · Cora | AUC | Hurt |
Surprising findings
- Neighbor pruning on head nodes also improves head-node tail-style performance, and the method improves head-node detection as well as tail nodes.
- Naive random edge completion degrades performance substantially (e.g., 78.28 AUC on Cora vs 92.83 for AD-GCL).
Limitations
- Focuses on homophilic graphs; effectiveness on heterophilic graphs is unconfirmed.
- Degree threshold K is fixed by prior work and the Pareto principle rather than adaptive.
- Lacks theoretical analysis of bounds for the method.
Experimental setup
Datasets: Cora Citeseer Pubmed Bitcoinotc BITotc BITalpha Reddit Tolokers
Benchmarks: Cora Citeseer Pubmed Bitcoinotc BITotc BITalpha Reddit Tolokers
Metrics: AUC AUPRC AP
Baselines: SCAN Radar ONE AEGIS GAAN CoLA ANEMONE AdONE GRADATE GAD-NR
Builds on: CoLA RWR-based subgraph contrast PC-GNN GCN
Closest papers by meaning
- Dual-discriminative Graph Neural Network for Imbalanced Graph-level Anomaly Detection (NeurIPS 2022 · similarity 0.89)
- Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.88)
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.87)
- Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection (ICML 2025 · similarity 0.87)
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.87)
- Rayleigh Quotient Graph Neural Networks for Graph-level Anomaly Detection (ICLR 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.