Research map / Graph learning / GNN architectures
Partitioning Message Passing 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 Partitioning Message Passing (PMP), which uses separate node-specific aggregation weights for fraud, benign, and unlabeled neighbors to handle label imbalance and heterophily in graph fraud detection without modifying graph structure.
Problem
Graph fraud detection with GNNs suffers from label imbalance and a heterophily-homophily mixture, and existing fixes that rewire or reweight graphs are costly and hard to scale.
Approach
Partition neighbor aggregation by label (fraud, benign, unlabeled) using distinct weight matrices, with unlabeled weights as a node-specific convex combination of the labeled ones and the weight matrices produced by generators conditioned on the center node; the paper shows this acts as a node-specific spectral filter.
Key results
- PMP reaches 93.97 AUC on Yelp (40% training) versus 90.57 for the best baseline GHRN, a gain of about 3.4 points.
- PMP reaches 99.62 AUC on T-Social (40% training), above BWGNN's 94.72.
- Ablation on Yelp shows GraphSAGE at 89.38 AUC rising to 93.15 with partitioning alone and 93.97 with the full PMP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace uniform neighbor aggregation with separate transformations for fraud and benign neighbors (ablation) | GraphSAGE · Yelp, Amazon, and T-Finance | AUC: 89.38 -> 93.15; 93.16 -> 96.33; 95.05 -> 95.92 | Helped |
| replace an independent unlabeled-neighbor matrix with an adaptive combination of the labeled-neighbor matrices (ablation) | GraphSAGE + partition · Yelp, Amazon, and T-Finance | AUC: 93.15 -> 93.51; 96.33 -> 97.61; 95.92 -> 96.86 | Helped |
| generate fraud and benign neighbor weight matrices as functions of the center node (ablation) | GraphSAGE + partition + adaptive combination · Yelp, Amazon, and T-Finance | AUC: 93.51 -> 93.97; 97.61 -> 97.57; 96.86 -> 97.10 | Mixed |
| use Partitioning Message Passing rather than the strongest listed baseline | PMP · Yelp (40% training) | AUC: 90.57 -> 93.97 | Helped |
| use Partitioning Message Passing rather than the strongest listed baseline | PMP · Amazon (40% training) | AUC: 97.07 -> 97.57 | Helped |
| use Partitioning Message Passing rather than the strongest listed baseline | PMP · T-Finance (40% training) | AUC: 95.84 -> 97.10 | Helped |
| use Partitioning Message Passing rather than the strongest listed baseline | PMP · T-Social (40% training) | AUC: 94.72 -> 99.62 | Helped |
| increase the number of message-passing layers beyond one (ablation) | PMP · Yelp, Amazon, and T-Finance | performance | Hurt |
Surprising findings
- Plain GraphSAGE is a strong baseline that outperforms several GFD-specific models, suggesting graph-structure augmentation may be unnecessary.
- On Amazon, PMP's F1-Macro (92.03) is below GHRN's (92.36), so PMP is not best on every metric.
Limitations
- Performance is sensitive to hidden dimension and number of layers, and the best setting varies by dataset.
- Theoretical analysis of the spectral interpretation assumes a regular graph and a simplified (linear, sum-aggregation) form.
Experimental setup
Datasets: YelpChi Amazon T-Finance T-Social Grab
Benchmarks: YelpChi Amazon T-Finance T-Social Grab
Metrics: AUC F1-Macro G-Mean
Baselines: GCN GraphSAGE GAT GPRGNN FAGCN Care-GNN PC-GNN H2-FDetector BWGNN GHRN GDN R-GCN
Builds on: GraphSAGE GCN BWGNN GHRN CARE-GNN
Closest papers by meaning
- Revisiting Graph-Based Fraud Detection in Sight of Heterophily and Spectrum (AAAI 2024 · similarity 0.93)
- Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.90)
- A Label-free Heterophily-guided Approach for Unsupervised Graph Fraud Detection (AAAI 2025 · similarity 0.90)
- DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph (AAAI 2024 · similarity 0.90)
- Provably Powerful Graph Neural Networks for Directed Multigraphs (AAAI 2024 · similarity 0.88)
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 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.