Research map / Graph learning / Graph self-supervised learning
DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph
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 DiG-In-GNN, a GNN fraud detector that uses multi-scale contrastive learning to generate discriminative guidance nodes against feature inconsistency and RL-based node-wise neighbor selection to counter structural inconsistency.
Problem
GNN-based fraud detection on multi-relation graphs is hurt by feature inconsistency and structural inconsistency caused by fraudsters' camouflage, which reduce the suspiciousness of fraud nodes after message passing.
Approach
Generate relation-wise guidance nodes with discriminative features via context-level and local-level contrastive learning, then use these guidance nodes to guide a policy-gradient RL neighbor selector that keeps only beneficial neighbors for message passing.
Key results
- On YelpChi, DiG-In-GNN reaches AUC 0.9357, F1-macro 0.8154 and GMean 0.8413, with the paper reporting up to 20.73% improvement over prior state of the art.
- On Amazon, DiG-In-GNN achieves AUC 0.9778 and F1-macro 0.9251, outperforming GTAN (0.9736 AUC) and other baselines.
- On T-Finance, DiG-In-GNN achieves AUC 0.9658 and GMean 0.8916, the best GMean among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DiG-In-GNN instead of the strongest listed baseline | DiG-In-GNN · YelpChi | AUC: 0.9108 -> 0.9357 | Helped |
| use DiG-In-GNN instead of the strongest listed baseline | DiG-In-GNN · Amazon | AUC: 0.9736 -> 0.9778 | Helped |
| use DiG-In-GNN instead of the strongest listed baseline | DiG-In-GNN · T-Finance | GMean: 0.8682 -> 0.8916 | Helped |
| add RL-based neighbor selection (ablation) | DiG-In-GNN · YelpChi | AUC: 0.9253 -> 0.9357 | Helped |
| add guidance node generation (ablation) | DiG-In-GNN · YelpChi | AUC: 0.9259 -> 0.9357 | Helped |
| remove neighbor aggregation and use guidance and raw features only (ablation) | DiG-In-GNN · YelpChi | AUC: 0.9357 -> 0.9345 | Hurt |
| add RL-based neighbor selection (ablation) | DiG-In-GNN · Amazon | AUC: 0.9736 -> 0.9778 | Helped |
| add guidance node generation (ablation) | DiG-In-GNN · Amazon | AUC: 0.9749 -> 0.9778 | Helped |
| remove neighbor aggregation and use guidance and raw features only (ablation) | DiG-In-GNN · Amazon | AUC: 0.9778 -> 0.9702 | Hurt |
| add RL-based neighbor selection (ablation) | DiG-In-GNN · T-Finance | GMean: 0.8731 -> 0.8916 | Helped |
| use context-level contrastive learning alone instead of the full two-scale guidance generation (ablation) | DiG-In-GNN · YelpChi | AUC: 0.9280 vs 0.9357 | Hurt |
| use local-level contrastive learning alone instead of the full two-scale guidance generation (ablation) | DiG-In-GNN · YelpChi | AUC: 0.9306 vs 0.9357 | Hurt |
Surprising findings
- Plain GCN outperforms several newer fraud detectors on T-Finance, which the authors attribute to that graph being more homophilous.
- DiG-In-GNN/AGG, which masks all neighbors and uses only raw and guidance node features, remains comparable to the full model, attributed to the high distinguishability of guidance nodes.
Limitations
- Evaluated only on three fraud datasets; the authors note generalization to multi-class node classification as future work.
- Neighbor sampling rates and reward tolerance values are set per relation and per dataset by hand, adding tuning burden.
- Per-relation separate parameter sets for guidance node generation scale with the number of relations.
Experimental setup
Datasets: YelpChi Amazon T-Finance
Benchmarks: YelpChi Amazon T-Finance
Metrics: AUC F1-macro GMean
Baselines: GCN GAT GraphSAGE Cluster-GCN GraphSAINT CARE-GNN PC-GNN RioGNN AO-GNN H2-FDetector GTAN GHRN
Builds on: GraphSAGE CARE-GNN RioGNN H2-FDetector GHRN
Closest papers by meaning
- Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.92)
- Revisiting Graph-Based Fraud Detection in Sight of Heterophily and Spectrum (AAAI 2024 · similarity 0.90)
- Partitioning Message Passing for Graph Fraud Detection (ICLR 2024 · similarity 0.90)
- Unveiling the Threat of Fraud Gangs to Graph Neural Networks: Multi-Target Graph Injection Attacks Against GNN-Based Fraud Detectors (AAAI 2025 · similarity 0.89)
- DGA-GNN: Dynamic Grouping Aggregation GNN for Fraud Detection (AAAI 2024 · similarity 0.88)
- Provably Powerful Graph Neural Networks for Directed Multigraphs (AAAI 2024 · similarity 0.88)
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.