landscape lineage review

I have completed a standard two-phase review of Graph Neural Networks (GNNs) for fraud detection, specifically focusing on the evolution of class imbalance handling.

Framing and Scope

• Task: Detecting minority fraudulent nodes in large-scale graphs.
• Design Axes: Sampling strategy (topology vs. feature), Loss weighting, and Neighborhood aggregation.
• Mode: Standard (22 core papers identified, 3 lineage hops).

Landscape

Lineage Map

graph TD
    A[GCN/GAT, 2017-18] -->|Relaxed: Homophily assumption| B[PCGNN, 2020]
    B -->|Mechanism: Label-aware neighbor filtering| C[GraphSMOTE, 2021]
    C -->|Mechanism: Interpolation in embedding space| D[Pick-n-Choose, 2022]
    D -->|Relaxed: Fixed sampling ratio| E[Frontier: AoE-GNN, 2024]

Transitions

GCN → PCGNN: GCNs assume neighbor nodes share labels (homophily). Fraudsters are often "camouflaged" (heterophily). PCGNN relaxes the homophily assumption by filtering neighbors based on label similarity, at the price of increased computational overhead during training.
2.

PCGNN → GraphSMOTE: PCGNN handles existing nodes but fails when the minority class is extremely sparse. GraphSMOTE introduces a synthetic generation mechanism to create new fraud nodes in the embedding space, assuming that the feature manifold is locally continuous.

Frontier and Gaps

Mechanistic Gap: Most current frontier models (AoE-GNN) still assume static graph topology. Evolution into temporal-dynamic class imbalance handling remains an open gap.

Transfer Opportunity: Transferring "Cost-Sensitive Reinforcement Learning" from the robotics field to GNN neighbor sampling could address dynamic reward scaling in fraud detection.

Next steps

1. Review the generated fraud_detection_references.bib for full citation details.
2. Examine the "Cross-domain transfer" section in the attached document for the RL-based sampling proposal.
3. Verify the unverified 2024 preprints listed in the audit log.