Research map / Graph learning / Graph self-supervised learning
A Label-free Heterophily-guided Approach for Unsupervised 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 HUGE, an unsupervised graph fraud detection method that estimates node heterophily label-free with a new HALO metric and trains a joint MLP-GNN with ranking and asymmetric alignment losses.
Problem
Graph fraud detection without labels is hard because fraudsters camouflage via heterophilic connections to benign users, and existing unsupervised GAD methods rely on homophily assumptions and iterative pseudo-labeling.
Approach
A label-free, attribute-based edge heterophily metric (HALO) satisfying boundedness, minimal agreement, monotonicity and equal attribute tolerance guides an MLP-based fraud scorer via a heterophily-order ranking loss, with an asymmetric KL alignment from a GNN branch to the MLP to inject structure without feature smoothing.
Key results
- On AmazonFull, HUGE reaches AUROC 0.8892 and AUPRC 0.7668, versus 0.6442 / 0.2188 for the next-best TAM.
- On Amazon, HUGE reaches AUROC 0.8516 and AUPRC 0.6672, versus 0.7126 / 0.2915 for TAM.
- Ablation: removing the asymmetric alignment loss drops Amazon AUROC from 0.8516 to 0.5614.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HUGE instead of the strongest baseline | Unsupervised GFD · Amazon, Facebook, YelpChi, AmazonFull, YelpChiFull | AUROC: 0.7126±0.0111 -> 0.8516±0.0029; 0.9539±0.0067 -> 0.9760±0.0008; 0.5473±0.0022 -> 0.6013±0.0076; 0.6442±0.0232 -> 0.8892± | Helped |
| Use HUGE instead of the strongest baseline | Unsupervised GFD · Reddit | AUROC: 0.6032±0.0057 -> 0.5906±0.0042 | Hurt |
| Use HUGE instead of the strongest baseline | Unsupervised GFD · Amazon, Facebook, Reddit, AmazonFull, YelpChiFull | AUPRC: 0.2915±0.0373 -> 0.6672±0.0077; 0.3630±0.0144 -> 0.3674±0.0091; 0.0465±0.0018 -> 0.0511±0.0022; 0.2396±0.0133 -> 0.7668± | Helped |
| Use HUGE instead of the strongest baseline | Unsupervised GFD · YelpChi | AUPRC: 0.0780±0.0014 -> 0.0708±0.0026 | Hurt |
| Replace HALO with Euclidean distance (ablation) | HUGE · Amazon, Facebook, AmazonFull, YelpChiFull | AUROC: 0.7128±0.0061; 0.9251±0.0074; 0.8746±0.0026; 0.5763±0.0064 | Helped |
| Replace HALO with cosine distance (ablation) | HUGE · Amazon, Facebook, AmazonFull, YelpChiFull | AUROC: 0.7018±0.0052; 0.9668±0.0024; 0.8735±0.0026; 0.5762±0.0063 | Helped |
| Replace HALO with attribute heterophily rate (ablation) | HUGE · Amazon, Facebook, AmazonFull, YelpChiFull | AUROC: 0.8171±0.0064; 0.9438±0.0029; 0.8622±0.0024; 0.5765±0.0064 | Helped |
| Add asymmetric alignment loss (ablation) | HUGE · Amazon, Facebook, AmazonFull, YelpChiFull | AUROC: 0.5614±0.0019 -> 0.8516±0.0029; 0.9487±0.0026 -> 0.9760±0.0008; 0.8746±0.0033 -> 0.8892±0.0026; 0.5295±0.0204 -> 0.5767± | Helped |
| Add GNN branch (ablation) | HUGE · Amazon, Facebook, AmazonFull, YelpChiFull | AUROC: 0.5548±0.0024 -> 0.8516±0.0029; 0.9401±0.0022 -> 0.9760±0.0008; 0.7005±0.0136 -> 0.8892±0.0026; 0.5533±0.0088 -> 0.5767± | Helped |
| Vary alignment weight α (ablation) | HUGE · Facebook, Reddit | AUROC | No clear effect |
| Set alignment weight α to 0 instead of the selected value (ablation) | HUGE · Amazon, Facebook | AUROC | Hurt |
| Vary alignment weight α (ablation) | HUGE · Amazon, YelpChi | AUROC | Mixed |
Surprising findings
- Replacing the alignment loss causes the largest drop on Amazon, while removing the GNN alone degrades YelpChiFull only slightly, and HUGE does not win every metric (e.g., AUROC on Reddit, AUPRC on YelpChi).
Limitations
- Node-level heterophily is a simple average of edge-level HALO; more principled node-level measures are left for future work.
- HALO only satisfies monotonicity under a relaxed constraint assuming constant attribute vector length.
- Performance is unstable on some datasets (e.g., Reddit AUROC is not best) and sensitive to alpha on Amazon and YelpChi.
Experimental setup
Datasets: Amazon Facebook Reddit YelpChi AmazonFull YelpChiFull
Benchmarks: Amazon Facebook Reddit YelpChi AmazonFull YelpChiFull
Metrics: AUROC AUPRC
Baselines: DOMINANT CoLA ANEMONE GRADATE ADA-GAD GADAM TAM
Builds on: CoLA DOMINANT
Closest papers by meaning
- Revisiting Graph-Based Fraud Detection in Sight of Heterophily and Spectrum (AAAI 2024 · similarity 0.91)
- Partitioning Message Passing for Graph Fraud Detection (ICLR 2024 · similarity 0.90)
- 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 0.87)
- Global Attribute-Association Pattern Aggregation for Graph Fraud Detection (AAAI 2025 · similarity 0.87)
- DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph (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.