Research map / Graph learning / Graph self-supervised learning

DGA-GNN: Dynamic Grouping Aggregation GNN for Fraud Detection

AAAI 2024 · Graph learning › Graph self-supervised learning · approach: Graph anomaly detection · 50 citations (OpenAlex)

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 DGA-GNN, a GNN for graph fraud detection that bins non-additive node attributes via decision-tree binning and dynamically groups neighbors into fraud/benign groups with feedback from the previous epoch, then aggregates hierarchically.

Problem

Graph-based fraud detection is hurt by non-additive node attributes (e.g., age, transaction frequency), which break mean aggregation, and by neighbor messages from fraudulent and benign nodes being mixed together, reducing their distinguishability.

Approach

Decision-tree binning encoding turns each attribute into one-hot bin vectors supervised by labels so that sum/mean aggregation is meaningful, and a feedback dynamic grouping partitions neighbors into predicted-fraud and predicted-benign groups using the previous epoch's predictions, followed by intra-group mean aggregation and GAT-style inter-group attention.

Key results

• DGA-GNN reaches AP 81.12 / AUC 94.22 on Elliptic and AP 92.80 / AUC 97.95 on YelpChi, improving over the next best method by 3%-16% across datasets per the paper.
• Ablation on T-Social and YelpChi shows removing the binning encoding or dynamic grouping lowers AP (e.g., YelpChi AP 92.80 full vs 78.73 without encoding and 80.72 without grouping).
• On YelpChi, the non-additive attribute subset with binning encoder reaches AP 90.02 vs 81.86 for the additive subset.

Measured effects reported in the paper

Surprising findings

• Random forest outperforms many GNN baselines such as GCN and GAT on several fraud datasets, suggesting decision-tree feature use is strong for fraud.

Limitations

• Grouping relies on a threshold z and number of bins k that are tuned on validation scores; performance peaks vary by dataset (k=4 vs k=32).
• Some heterophilic baselines (AMNet, H2-FDetector) run out of memory on large datasets, so comparisons are incomplete there.
• Future work noted: integrating decision trees with GNNs and exploring heterophily further.

Experimental setup

Datasets: Elliptic T-Finance T-Social YelpChi Amazon

Benchmarks: Elliptic T-Finance T-Social YelpChi Amazon

Metrics: AP AUC

Baselines: MLP RF GCN GAT AMNet BWGNN GHRN CARE-GNN PC-GNN RioGNN H2-FDetector GAGA

Builds on: GAGA GAT BWGNN CARE-GNN

Closest papers by meaning

• Global Attribute-Association Pattern Aggregation for Graph Fraud Detection (AAAI 2025 · similarity 0.89)
• Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.89)
• DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph (AAAI 2024 · similarity 0.88)
• Unveiling the Threat of Fraud Gangs to Graph Neural Networks: Multi-Target Graph Injection Attacks Against GNN-Based Fraud Detectors (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)