Research map / Graph learning / Graph self-supervised learning

Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection

AAAI 2025 · Graph learning › Graph self-supervised learning · approach: Graph anomaly detection · 3 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 DCL-GFD, a GNN for graph-based fraud detection that estimates node abnormality via node-subgraph contrastive learning and uses the resulting probabilities to dynamically split neighbor aggregation into benign-like and fraud-like channels.

Problem

Graph-based fraud detection suffers from homophily-heterophily mixture and severe class imbalance, causing standard homophily-based GNNs to lose local abnormal signals and under-represent rare fraud nodes.

Approach

Estimate each node's abnormal probability by cosine similarity between the node embedding and its RWR subgraph embedding (trained with a label-aware InfoNCE-style loss), then use that probability to weight neighbor aggregation separately into benign-like and fraud-like embeddings, jointly optimized end to end.

Key results

• DCL-GFD reaches 98.99% AUC on Weibo (vs. 97.22% best baseline BWGNN in table, text reports up to 2.03% gain), and 93.16% Macro-F1 on Amazon versus 92.43% for CONSISGAD.
• DCL-GFD outperforms all baselines on 3 of 4 datasets; on T-Finance it ranks second behind PMP/CONSISGAD in AUC (96.58 vs 96.87 for CONSISGAD).
• Gains over GraphSage are 13.06% AUC on Weibo and 15.76% AUC on YelpChi.

Measured effects reported in the paper

Surprising findings

• On T-Finance the method performs second best, which the authors attribute to limited initial feature information hurting the node abnormality estimation.

Limitations

• Performance on T-Finance is only second best, attributed to limited node features weakening abnormality estimation.
• Relies on partially observed training labels to form the contrastive loss, so it depends on label availability.
• Sensitive to subgraph sampling hyperparameters (step size k and restart probability r).

Experimental setup

Datasets: Weibo Amazon YelpChi T-Finance

Benchmarks: Weibo Amazon YelpChi T-Finance

Metrics: AUC Macro-F1

Baselines: GCN GAT GraphSage AMNet BWGNN GHRN CARE-GNN PC-GNN GDN PMP CONSISGAD

Builds on: GCN InfoNCE random walk with restart CONSISGAD PMP

Closest papers by meaning

• DiG-In-GNN: Discriminative Feature Guided GNN-Based Fraud Detector against Inconsistencies in Multi-Relation Fraud Graph (AAAI 2024 · similarity 0.92)
• 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)
• DGA-GNN: Dynamic Grouping Aggregation GNN for Fraud Detection (AAAI 2024 · similarity 0.89)
• Context-aware Graph Neural Network for Graph-based Fraud Detection with Extremely Limited Labels (AAAI 2025 · similarity 0.88)
• Global Attribute-Association Pattern Aggregation for Graph Fraud Detection (AAAI 2025 · similarity 0.88)