Federated Graph Learning

Each client holds a local subgraph representing its own patient population, drug interactions, or molecular structures. The global model learns to generalize across these isolated subgraphs by sharing only GNN model parameters—never raw adjacency matrices or node features. This enables cross-hospital knowledge graphs to grow without violating privacy.

• Missing neighbor handling: Nodes at subgraph boundaries lack visibility into cross-institution edges, requiring specialized neighborhood sampling
• Overlapping nodes: When the same entity exists across silos, entity alignment techniques resolve identity without sharing identifiers
• Global structure inference: The aggregated model implicitly learns topological patterns present across the entire distributed graph

GNNs operate through neighborhood aggregation—each node updates its representation by collecting features from its neighbors. In federated settings, this creates a tension: message passing requires neighbor information, but privacy prohibits sharing it. Solutions include:

• Localized message passing: Each client runs full GNN layers on its subgraph, treating boundary nodes as leaf nodes
• Differential privacy on gradients: Noise is injected into shared model updates to prevent reconstruction of local graph structure
• Secure aggregation: Cryptographic protocols ensure the central server can only compute the sum of client gradients, never individual contributions
• Split GNN architectures: The model is partitioned so clients only share intermediate embeddings, not raw node features

Unlike standard federated learning where data is assumed IID, graph data across institutions is inherently non-IID in multiple dimensions:

• Structural heterogeneity: One hospital's patient graph may be dense with many co-morbidity edges, while another's is sparse
• Feature heterogeneity: Different institutions collect different clinical attributes for the same entity types
• Label heterogeneity: Diagnostic coding standards vary across healthcare systems

Federated graph learning must address all three simultaneously through personalized GNN layers, feature alignment modules, and multi-task learning heads that adapt the global model to each site's specific graph distribution.

Training GNNs on massive graphs requires neighborhood sampling to create manageable mini-batches. In federated settings, sampling must respect data locality:

• Client-side sampling: Each institution independently samples subgraphs from its local data for each training round
• Importance-based node selection: Nodes with high betweenness centrality or boundary positions are prioritized to preserve global structure
• Temporal sampling: For dynamic graphs like patient trajectories, sampling windows must align across institutions to maintain causal consistency
• Communication-efficient sampling: The number of sampled nodes per round is tuned to balance model accuracy against bandwidth constraints

Graph Attention Networks (GATs) learn to weight the importance of different neighbors during message passing. In federated settings, attention coefficients present unique challenges:

• Cross-silo attention blindness: A node cannot attend to neighbors residing on a different client's subgraph
• Attention alignment: The global model must learn attention patterns that generalize across heterogeneous local graph structures
• Privacy-preserving attention: Attention weights could leak information about local graph topology; techniques like attention dropout and gradient clipping mitigate this risk
• Federated attention pooling: Global readout functions that aggregate node embeddings into graph-level representations must operate without seeing the full graph structure

Federated graph learning enables pharmaceutical companies and research hospitals to collaboratively train models on molecular interaction graphs without revealing proprietary compound structures:

• Federated molecular property prediction: Each institution trains on its private library of compounds, sharing only GNN weights
• Cross-institution protein-protein interaction networks: Hospitals contribute patient-specific PPI subgraphs to build comprehensive interaction maps
• Adverse drug reaction graphs: Pharmacovigilance data from multiple healthcare systems is combined to detect rare side effect patterns
• De novo drug design: Generative graph models trained federatedly produce novel molecular structures informed by diverse, private datasets

The foundational deep learning architecture that operates directly on graph-structured data. GNNs learn node representations by iteratively aggregating feature information from neighboring nodes—a process known as message passing. In healthcare, GNNs model patient-drug interactions, protein-protein interaction networks, and disease ontologies. Key variants include Graph Convolutional Networks (GCNs) for transductive learning, GraphSAGE for inductive node embedding, and Graph Attention Networks (GATs) that learn to weight the importance of different neighbors. Federated Graph Learning extends these architectures to train across distributed subgraphs without centralizing sensitive relational data.

The process of dividing a large global graph into smaller, non-overlapping subgraphs distributed across federated clients. Effective partitioning is critical because cross-edge dependencies—edges that connect nodes assigned to different clients—create information loss. Strategies include METIS for balanced partitioning, random vertex-cut for power-law graphs, and semantic partitioning where subgraphs align with natural institutional boundaries (e.g., each hospital's patient population). Poor partitioning leads to neighborhood truncation, where a node's local computation misses critical distant neighbors, degrading GNN performance.

Techniques to compensate for missing cross-subgraph edges during federated training. Since clients cannot share raw adjacency data, methods include virtual node injection where a learnable embedding represents the global context, neighborhood expansion using differentially private node degree statistics, and graph structure proxy generation via federated generative models. A prominent approach is FedSage+, which trains a local graph generator to predict missing neighbors and their features, then shares these synthetic nodes across clients to reconstruct the global graph topology without exposing real patient connections.

Extends the attention mechanism to decentralized graph settings where attention coefficients must be computed across client boundaries. Federated Graph Attention Networks (FedGAT) enable each client to compute local attention scores while a central server aggregates cross-client attention patterns. Key challenges include attention score privacy—raw attention weights can leak structural information—and cross-subgraph attention alignment where nodes in different clients attend to the same global context. Solutions involve secure aggregation of attention matrices and federated multi-head attention where different heads specialize in local versus global structural patterns.

The most common federated graph learning task, where the goal is to predict labels for nodes whose features and edges span multiple clients. Each client holds a subgraph with partially labeled nodes and trains a local GNN. The global model must aggregate node embeddings without direct access to cross-client edges. Semi-supervised scenarios dominate in healthcare—e.g., predicting disease phenotypes where only a fraction of patients have confirmed diagnoses. Performance metrics focus on macro F1-score across clients to ensure fairness, as label distribution skew can cause some institutions to suffer degraded accuracy.

Predicting missing or future edges in a distributed graph without centralizing adjacency information. Critical applications include drug-drug interaction prediction across hospital formularies and patient similarity network completion for cohort discovery. The challenge is that candidate edges often connect nodes in different clients, requiring cross-client node pair scoring. Approaches include federated node embedding alignment where clients project embeddings into a shared latent space, and split-link prediction where each client computes partial scores that are securely aggregated. Hits@K and Mean Reciprocal Rank (MRR) are standard evaluation metrics.