Reconstructing Cross-Silo Connections in Federated Graph Learning

This direction is merged into another direction: graph partitioning trade-offs. Its last version stays readable and citable here.

Version 1 of 122 papers37 conceptsfrozen 2026-10-04

In standard Graph Neural Networks (GNNs), a node’s representation is defined by its neighborhood. By aggregating features from adjacent nodes, GNNs capture the structural context necessary for tasks like node classification or link prediction. However, when a graph is distributed across multiple private silos—as is the case in Federated Learning (FL)—this fundamental mechanism breaks down.

The core challenge in this research direction is the "shattered graph" problem. When a global graph is partitioned into local subgraphs, the edges that once connected nodes across different silos are severed. This leaves local models with a fragmented view of the world, missing the "cross-silo" neighbors that provide critical structural signals.

The Missing Neighbor Problem

When data is siloed, nodes located at the boundaries of a partition lose their external connections. As Han Yu, Pengwei Xing, and their colleagues note in Federated Graph Neural Networks: Overview, Techniques and Challenges, this "messiness" prevents GNNs from accounting for global structural dependencies. Without these missing edges, the message-passing process is truncated, leading to suboptimal embeddings.

To solve this, researchers have moved beyond simple model aggregation (like FedAvg) toward structural reconstruction: the art of "mending" or "inpainting" the missing pieces of the graph without actually exchanging raw private data.

Generative Mending and Network Inpainting

One powerful intuition is that if we cannot see the missing neighbors, we can learn to "hallucinate" them. In Subgraph Federated Learning with Missing Neighbor Generation, Carl Yang and Ke Zhang introduce FedSage+, which uses a generative model to synthesize representative neighbor nodes and edges. By training a generator to produce synthetic structural context, the local GNN can perform message passing as if the cross-silo connections were still there.

A similar approach is taken in the medical domain by Liang Peng and Xiaoxiao Li. Their framework, FedNI, treats the missing data as a "network inpainting" problem. Using a Generative Adversarial Network (GAN), FedNI: Federated Graph Learning With Network Inpainting for Population-Based Disease Prediction reconstructs missing nodes and edges in local graphs. This allows local models to leverage global population patterns that would otherwise be lost to data siloing.

Global Priors as Structural Scaffolding

If generative models "guess" the missing neighbors, another strategy uses the central server to provide a "scaffold" that guides local reconstruction. The server can aggregate non-sensitive, high-level structural statistics to help clients understand where their local subgraphs fit into the global topology.

For example, in Federated Node-Level Clustering Network with Cross-Subgraph Link Mending, Jingxin Liu and Wenxuan Tu use "clustering priors" to estimate the affinity between nodes across different partitions. By mending these links on the server, they restore the structural similarity of the original graph, which significantly boosts unsupervised clustering performance.

Similarly, Guoqing Chao and Zhenghao Zhang propose FIMCFG, a framework where a "globally fused graph" acts as a guide. As detailed in Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance, local clients repair their adjacency matrices by substituting missing rows with information derived from this global structure. This allows nodes with missing data views to estimate their relationships based on the collective connectivity patterns of the entire network.

Restoring Global Semantics

The ultimate goal of these techniques is to restore "global semantics"—the high-order relationships that emerge only when the graph is viewed as a whole. Yang Cao and Chuan Shi demonstrate this in Federated Heterogeneous Graph Neural Network for Privacy-preserving Recommendation. Their FedHGNN framework "repairs" the broken semantics of multi-typed relationships in recommendation systems, showing that even sparse, voluntarily shared data can be made robust through structural reconstruction.

Whether through GAN-based inpainting, synthetic neighbor generation, or global probability propagation (as seen in the MV-FGAD work by Junyi Yan and KE LIANG), the intuition remains the same: the loss of connectivity in federated settings is a recoverable signal. By treating the missing edges not as a permanent loss, but as a reconstruction task, Federated GNNs can achieve the performance of centralized models while maintaining the privacy of the underlying data.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_reconstructing_cross_silo_connections_in_b6f2477c,
  title = {Reconstructing Cross-Silo Connections in Federated Graph Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reconstructing-cross-silo-connections-in-federated-graph-learning/rdr_b6f2477c977b4682b9c7c70dbfa1dc6a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_reconstructing_cross_silo_connections_in_b6f2477c_v1,
  title = {Reconstructing Cross-Silo Connections in Federated Graph Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/reconstructing-cross-silo-connections-in-federated-graph-learning/rdr_b6f2477c977b4682b9c7c70dbfa1dc6a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}