Research map / Graph learning / Graph self-supervised learning
Global Attribute-Association Pattern Aggregation for Graph 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 GAAP, an end-to-end graph fraud detector that learns dynamic bin embeddings for attributes, aggregates them with GraphSAGE over neighbors, and globally aggregates node patterns via cross-attention against cached historical embeddings.
Problem
Graph fraud detection where random forests miss graph associations and GNNs incorrectly fuse heterogeneous attribute values, so fraud patterns are poorly captured.
Approach
Adaptively split each attribute into learnable bins (dynamic binning embedding with PLE-style encoding) so attributes are not averaged together, pass these attribute-pattern features through a GNN to capture association patterns, and globally aggregate node representations with cross-attention to cached historical embeddings at O(N·M) cost.
Key results
- GAAP achieves average Rec@K of 70.88% across 7 datasets, the best among 24 baselines (DGA-GNN 69.40%, XGBGraph 68.56%).
- On YelpChi, GAAP reaches 88.54% Rec@K versus 84.23% for DGA-GNN, an improvement of about 4.3 points.
- Ablations show removing the DyBEM module drops YelpChi Rec@K from 87.51% to 47.62%, and replacing cross-attention global aggregation with graph partitioning reduces YelpChi Rec@K to 68.19%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add dynamic binning embedding (DyBEM) (ablation) | GAAP · YelpChi | Rec@K: 47.62 -> 87.51 | Helped |
| Add dynamic binning embedding (DyBEM) (ablation) | GAAP · T-Finance | Rec@K: 79.07 -> 85.22 | Helped |
| Add dynamic binning embedding (DyBEM) (ablation) | GAAP · T-Social | Rec@K: 75.82 -> 97.05 | Helped |
| Use dynamic binning instead of static one-hot encoding (ablation) | GAAP · YelpChi | Rec@K: 75.33 -> 87.51 | Helped |
| Use dynamic binning instead of static one-hot encoding (ablation) | GAAP · T-Finance | Rec@K: 84.49 -> 85.22 | Helped |
| Use dynamic binning instead of static one-hot encoding (ablation) | GAAP · T-Social | Rec@K: 95.52 -> 97.05 | Helped |
| Add the message-passing GNN module (ablation) | GAAP · YelpChi | Rec@K: 61.62 -> 87.51 | Helped |
| Add the message-passing GNN module (ablation) | GAAP · T-Finance | Rec@K: 72.63 -> 85.22 | Helped |
| Add the message-passing GNN module (ablation) | GAAP · T-Social | Rec@K: 23.32 -> 97.05 | Helped |
| Add cross-attention global aggregation (ablation) | GAAP · YelpChi | Rec@K: 86.29 -> 87.51 | Helped |
| Add cross-attention global aggregation (ablation) | GAAP · T-Finance | Rec@K: 84.12 -> 85.22 | Helped |
| Add cross-attention global aggregation (ablation) | GAAP · T-Social | Rec@K: 96.85 -> 97.05 | Helped |
Surprising findings
- Replacing the global cross-attention with graph-partition-based aggregation performs worse than no global aggregation on some datasets, attributed to loss of randomness in node connections.
- Batch size had no significant impact on overall performance, per the authors' analysis.
Limitations
- Performance on Amazon is only second-best, which the authors attribute to noisy graph structure.
- Dynamic binning is only validated on fraud graph tabular data; applicability to CV and NLP is suggested but not tested.
- Results on DGraph-Fin are weaker than some baselines (e.g., BGNN, PNA-level scores in table), with GAAP at 7.73 vs. best 7.70 for BGNN in the Rec@K column.
Experimental setup
Datasets: Amazon YelpChi T-Finance Elliptic Tolokers DGraph-Fin T-Social
Benchmarks: Amazon YelpChi T-Finance Elliptic Tolokers DGraph-Fin T-Social
Metrics: Rec@K AUROC AUPRC
Baselines: MLP KNN Random Forest XGBoost Neighborhood Averaging GCN GraphSAGE GAT GIN SGC GT BGNN
Builds on: GraphSAGE piecewise linear encoding DGA-GNN PMP GADBench
Closest papers by meaning
- DGA-GNN: Dynamic Grouping Aggregation GNN for Fraud Detection (AAAI 2024 · similarity 0.89)
- Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · 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)
- A Label-free Heterophily-guided Approach for Unsupervised Graph Fraud Detection (AAAI 2025 · similarity 0.87)
- Partitioning Message Passing for Graph Fraud Detection (ICLR 2024 · similarity 0.86)
- Revisiting Graph-Based Fraud Detection in Sight of Heterophily and Spectrum (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.