Research map / Graph learning / Graph self-supervised learning
Context-aware Graph Neural Network for Graph-based Fraud Detection with Extremely Limited Labels
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 CGNN, a context-aware GNN for graph fraud detection with 1 or 5 labeled fraud nodes, using category semantic decomposition, denoising attention, and subspace feature perturbation with entropy-based consistency regularization.
Problem
Graph-based fraud detection when only extremely few labeled (fraud and normal) nodes are available, with most nodes unlabeled ("gray samples") and fraud camouflage/heterophily complicating aggregation.
Approach
Decompose each node's features into normal and abnormal category subspaces with a contrastive loss, use attention to adaptively combine homophilic and heterophilic neighbor messages, and apply subspace-wise feature perturbation to unlabeled nodes with a consistency loss, giving semantic guidance to gray samples.
Key results
- With 1 labeled fraud node, CGNN reaches 85.56% Macro-F1 on Amazon versus 80.08% for ConsisGAD (next best in that column).
- CGNN achieves 90.08% Macro-F1 on Elliptic with 1 fraud label and 86.36% on T-Finance with 1 label, the best Macro-F1 on all four datasets.
- Ablation shows removing category semantic decomposition, consistency training, or attention each degrades performance, supporting the modules.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CGNN instead of ConsisGAD | Graph-based fraud detection · Amazon (1 labeled abnormal node) | Macro-F1: 80.08±13.70 -> 85.56±8.49 | Helped |
| Use CGNN instead of GHRN | Graph-based fraud detection · Amazon (5 labeled abnormal nodes) | Macro-F1: 89.99±1.15 -> 90.08±0.72 | Helped |
| Use CGNN instead of BSL | Graph-based fraud detection · Elliptic (1 labeled abnormal node) | Macro-F1: 65.13±9.63 -> 76.26±7.07 | Helped |
| Use CGNN instead of ConsisGAD | Graph-based fraud detection · Elliptic (5 labeled abnormal nodes) | Macro-F1: 75.64±11.1 -> 82.79±4.48 | Helped |
| Use CGNN instead of ConsisGAD | Graph-based fraud detection · T-Finance (1 labeled abnormal node) | Macro-F1: 80.10±6.63 -> 86.36±3.59 | Helped |
| Use CGNN instead of ConsisGAD | Graph-based fraud detection · T-Finance (5 labeled abnormal nodes) | Macro-F1: 88.37±2.33 -> 90.10±1.62 | Helped |
| Use CGNN instead of ConsisGAD | Graph-based fraud detection · YelpChi (1 labeled abnormal node) | Macro-F1: 51.90±3.05 -> 52.71±1.13 | Helped |
| Use CGNN instead of ConsisGAD | Graph-based fraud detection · YelpChi (5 labeled abnormal nodes) | Macro-F1: 55.05±1.96 -> 56.08±1.86 | Helped |
| Add category semantic decomposition (CSD) (ablation) | CGNN · Four fraud-detection datasets | Macro-F1 | Helped |
| Add consistency training for unlabeled data (ablation) | CGNN · Four fraud-detection datasets | Macro-F1 | Helped |
| Add the denoising attention mechanism (ablation) | CGNN · Four fraud-detection datasets | Macro-F1 | Helped |
| Vary the category-semantic temperature coefficient τ (ablation) | CGNN · Amazon and Elliptic | Macro-F1 | No clear effect |
Surprising findings
- On YelpChi, all methods perform near chance (about 50% recall), and CGNN shows only comparable performance there.
Limitations
- Gains on YelpChi are small or comparable to baselines.
- Performance sensitive to temperature and loss weights in some settings, requiring tuning.
- Evaluated only on four datasets with label-scarce splits of 1 or 5 fraud nodes.
Experimental setup
Datasets: Amazon YelpChi Elliptic T-Finance
Benchmarks: Amazon YelpChi Elliptic T-Finance
Metrics: Macro-F1 Macro-Recall Macro-Precision
Baselines: GCN GAT GraphSAGE CARE-GNN FRAUDRE PC-GNN AMNet BWGNN GHRN GTAN Semi-GNN BSL
Builds on: ConsisGAD BWGNN CARE-GNN FixMatch Supervised contrastive learning
Closest papers by meaning
- Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.88)
- Barely Supervised Learning for Graph-Based Fraud Detection (AAAI 2024 · similarity 0.88)
- DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph (AAAI 2024 · similarity 0.87)
- Partitioning Message Passing for Graph Fraud Detection (ICLR 2024 · similarity 0.87)
- Revisiting Graph-Based Fraud Detection in Sight of Heterophily and Spectrum (AAAI 2024 · similarity 0.86)
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.86)
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.