Graph Federated Learning: Necessity, Classification, Algorithms, Platform Architecture, and Financial Applications

Graph federated learning (GFL) combines federated learning and graph neural networks to enable collaborative model training across multiple parties while preserving data privacy, a concept first introduced by Google in 2017 and later merged with graph learning techniques.

Necessity: GFL addresses the need for cross‑institutional data collaboration in regulated domains such as finance, where data cannot be centrally aggregated but complementary features and graph structures can improve tasks like anti‑money‑laundering and fraud detection.

Classification: GFL can be categorized by federated learning type (cross‑device, cross‑structure, horizontal, vertical, hybrid), by architecture (client‑server, peer‑to‑peer, committee), by handling of Non‑IID data, and by model purpose (node classification, graph classification, link prediction). Architectural choices affect privacy guarantees and communication efficiency.

Classic Algorithms: Representative models include FedGraphNN (using GCN and SecAggregate), FedGraph with advanced sampling and reinforcement‑learning‑based weight adjustment, VFGNN for vertical federated scenarios, and DPA‑SGD for decentralized optimization. These algorithms demonstrate improvements in accuracy and training speed over centralized baselines.

Platform Architecture: A GFL platform typically consists of a federated learning core deployed via Docker containers, communication layers (gRPC, Kafka), and graph data handling modules (storage, preprocessing, alignment, pruning). Open‑source examples such as FedGraphNN illustrate integration of GNN kernels and privacy‑preserving protocols.

Financial Applications: In anti‑money‑laundering, horizontal GFL enables banks to jointly model transaction graphs; in fraud detection, vertical GFL merges heterogeneous data (e.g., insurance, vehicle, personal networks) to enhance risk prediction. These use cases highlight the benefits of cross‑entity graph collaboration.

Future Outlook: Challenges remain in scaling to massive dynamic graphs, optimizing for Non‑IID distributions, handling heterogeneous and temporal graph structures, and achieving decentralized yet privacy‑secure communication. Further research is needed on distributed graph storage, automated hyper‑parameter tuning, and robust consensus mechanisms.

Q&A Highlights: CS architecture is preferred when a trusted coordinator exists, while P2P or committee structures suit large, mutually distrustful participant sets; sampling for Non‑IID mitigation often occurs at hidden‑layer levels; horizontal GFL fits similar‑type institutions, whereas vertical GFL addresses feature‑complementary cross‑industry collaborations.