Research map / Graph learning / Graph self-supervised learning
Barely Supervised Learning for Graph-Based Fraud 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 BSL, a barely supervised graph fraud detector that combines edge-type feature disentanglement on a few labeled nodes with consistency-based weak/strong augmentation on unlabeled nodes.
Problem
Graph-based fraud detection methods assume large labeled training sets, but in practice annotated fraud nodes are very scarce (about 1% or fewer), and camouflage makes detection harder under these conditions.
Approach
Disentangle node embeddings into three subspaces (normal-abnormal, abnormal-abnormal, normal-normal) using edge classification and attention, then build consistency and inconsistency losses on unlabeled nodes by swapping subspaces from labeled normal and abnormal exemplars to form weak and strong augmentations.
Key results
- On Amazon with 1% labels, BSL reaches 91.42 macro-F1 versus 88.59 for PC-GNN and 85.28 for GTAN.
- On Elliptic with 1% labels, BSL reaches 87.92 macro-F1 versus 82.99 for BWGNN.
- On T-Social with 0.01% labels, BSL reaches 75.67 macro-F1 versus 72.14 for BWGNN; BSL achieves best macro-F1 on all five datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BSL instead of PC-GNN | Fraud detection · Amazon (1%) | macro-F1: 88.59 -> 91.42 | Helped |
| Use BSL instead of BWGNN | Fraud detection · Yelp (1%) | macro-F1: 61.92 -> 62.83 | Helped |
| Use BSL instead of BWGNN | Fraud detection · Elliptic (1%) | macro-F1: 82.99 -> 87.92 | Helped |
| Use BSL instead of GTAN | Fraud detection · T-Finance (1%) | macro-F1: 80.08 -> 86.10 | Helped |
| Use BSL instead of BWGNN | Fraud detection · T-Social (0.01%) | macro-F1: 72.14 -> 75.67 | Helped |
| Add the barely supervised learning module (ablation) | BSL · Five datasets | detection performance | Helped |
| Add the feature disentanglement module (ablation) | BSL · Five datasets | detection performance | Helped |
Surprising findings
- GTAN degrades sharply on Yelp (macro-F1 45.75) and T-Social under extreme label scarcity, suggesting label-propagation-based methods are fragile in this regime.
Limitations
- AMNet runs out of memory on T-Social, showing scalability limits of the baseline, while the proposed method's scalability is not analyzed in depth.
- Validation set is not available, so model selection uses training loss, which is a weak selection signal.
- Strong/weak augmentation design relies on heuristic assumptions (binary fraud setting, NA subspace has limited category information).
Experimental setup
Datasets: Amazon Yelp Elliptic T-Finance T-Social
Benchmarks: Amazon Yelp Elliptic T-Finance T-Social
Metrics: macro-precision macro-recall macro-F1
Baselines: GCN GraphSage GAT Semi-GNN CARE-GNN FRAUDRE PC-GNN AMNet BWGNN GTAN
Builds on: FixMatch GCN Semi-GNN BWGNN CARE-GNN
Closest papers by meaning
- Revisiting Graph-Based Fraud Detection in Sight of Heterophily and Spectrum (AAAI 2024 · similarity 0.89)
- Context-aware Graph Neural Network for Graph-based Fraud Detection with Extremely Limited Labels (AAAI 2025 · similarity 0.88)
- A Label-free Heterophily-guided Approach for Unsupervised Graph Fraud Detection (AAAI 2025 · similarity 0.87)
- Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.87)
- UniFORM: Towards Unified Framework for Anomaly Detection on Graphs (AAAI 2025 · similarity 0.86)
- DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph (AAAI 2024 · 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.