Healing Fragmented Graphs for Federated Learning

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2152 papers191 conceptsfrozen 2026-10-04

In traditional machine learning, data points are often treated as independent. If you split a dataset of images between two servers, each server still has complete images. But graphs are different. In a graph, the "data" lives in the relationships between points. When you partition a large social network or a financial transaction web across different silos for privacy, you don't just distribute the nodes—you sever the edges.

This phenomenon, often called structural fragmentation, is the central challenge of Subgraph Federated Learning (SFL). As Huanding Zhang and Chao Wu argue in Federated Graph Learning - A Position Paper, partitioning a global graph into local subgraphs restricts the "receptive field" of Graph Neural Networks (GNNs). A node's representation depends on its neighbors, but in a federated setting, those neighbors might be locked in a different silo.

The Missing Neighbor Problem

When edges are cut, GNNs lose their ability to perform message passing across the full global topology. This leads to several failures:

  1. Broken High-Order Patterns: Recommendation systems, like those discussed in FedGNN, rely on "multi-hop" relationships (e.g., a friend of a friend liked this item). If the "friend" is on a different device, the connection is invisible.
  2. Label Signal Disruption: In semi-supervised learning, labels propagate through edges. Zihan Tan and He Li show in $S^2$FGL that as fragmentation increases, the influence of labeled nodes on the rest of the network drops sharply.
  3. Incomplete Context: For security tasks like intrusion detection, Jiacen Xu notes in ENTENTE that lateral movement attacks often span multiple administrative domains. A local view sees only isolated events, missing the broader attack graph.

Mending the Graph with Generative Models

One of the most intuitive solutions is to "hallucinate" what is missing. If we can't see the neighbors in other silos, can we predict what they would look like?

Carl Yang and colleagues pioneered this with FedSage+. They introduced a generative module that "mends" local subgraphs by creating synthetic neighbors. These synthetic nodes act as proxies for the missing cross-silo connections, allowing the GNN to perform message passing as if the graph were whole.

A similar approach is taken in FedNI by Liang Peng, which uses a GAN-based "network inpainting" mechanism. By reconstructing the missing links before training the classifier, the model "heals" the fragmented population graph, significantly improving disease prediction accuracy.

Decoupling Structure from Features

If we cannot share raw data, perhaps we can share the "shape" of the data. Johan Östman and Alexandre Graell i Amat propose FedStruct, which decouples the graph structure from node features.

They use a method called Hop2Vec to create task-dependent embeddings that capture structural context without exposing sensitive node attributes. This allows the model to understand that two nodes are connected across silos without ever knowing who or what those nodes are. By focusing on the "skeleton" of the global graph, FedStruct achieves accuracy nearly identical to centralized models.

Global Workspaces and Clustering Priors

Another emerging strategy is to create a "middle ground" where silos can synchronize. Jiawei Sun introduces the "Graph Global Workspace" in InterpGNN. This acts as a shared memory buffer, allowing distant or out-of-batch nodes to communicate structural signals.

Similarly, FedNCN by Jingxin Liu uses "clustering prior knowledge" on a central server. Instead of trying to reconstruct every individual edge, the server identifies broad community structures. These priors are then sent back to the clients to "mend" the local view, ensuring that local updates are consistent with the global community topology.

Why It Works: Restoring the Receptive Field

The success of these methods—whether through generative inpainting, structural decoupling, or global workspaces—stems from a single principle: restoring the effective receptive field.

GNNs are powerful because they leverage the "wisdom of the crowd" (the neighborhood). Fragmentation silences that crowd. By "healing" the graph, these frameworks allow the model to see beyond its immediate silo. As Guowei Wu demonstrates with GNN4FR, when you successfully synchronize these high-order interactions, federated training becomes "lossless"—producing the exact same accuracy as if all the data were sitting on a single, centralized server.

Papers in this version

Showing 60 of 152 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 140 papers
  2. Version 1 · 2026-10-04 · 152 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_graph_partitioning_trade_offs_cd7f690f,
  title = {graph partitioning trade-offs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/graph-partitioning-trade-offs/rdr_cd7f690f728e4dbcab0eae1c089eb956},
  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_healing_fragmented_graphs_for_federated__cd7f690f_v1,
  title = {Healing Fragmented Graphs for Federated Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/healing-fragmented-graphs-for-federated-learning/rdr_cd7f690f728e4dbcab0eae1c089eb956/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}