SpreadFGL: Edge-Client Collaborative Federated Graph Learning with Adaptive Neighbor Generation

Luying Zhong, Yueyang Pi, Zheyi Chen, Zhengxin Yu, Wang Miao, Xing Chen, Geyong Min

Abstract

Federated Graph Learning (FGL) enables training Graph Neural Networks (GNNs) across distributed clients without sharing raw data. However, FGL faces two critical challenges: the "data island" effect due to missing inter-client links and the scalability bottleneck of centralized server architectures. We propose SpreadFGL, an edge-client collaborative framework designed to mend these data islands. SpreadFGL utilizes an adaptive graph imputation generator at the edge layer to reconstruct missing links and a distributed multi-edge architecture to enhance scalability. Our evaluation shows that SpreadFGL achieves up to a 12.78% accuracy improvement over state-of-the-art methods with significantly faster convergence.

Introduction

Graph Neural Networks (GNNs) have revolutionized tasks such as social network analysis and citation classification. In practice, graph data is often siloed across multiple clients (e.g., different organizations or devices) due to privacy regulations. Federated Graph Learning (FGL) addresses this by allowing collective training without raw data exchange.

Despite its potential, FGL is hindered by two major issues. First, partitioning a global graph into local subgraphs causes the loss of inter-client links, preventing GNNs from aggregating information from multi-hop neighbors across client boundaries. Second, traditional FGL relies on a single central server, which becomes a communication and computation bottleneck as the network scales.

Methodology: The SpreadFGL Framework

SpreadFGL introduces a collaborative architecture between clients and multiple edge servers to address these challenges.

Adaptive Graph Imputation Generator

To overcome the missing link problem, SpreadFGL employs an edge-layer imputation generator. Instead of sharing raw data, clients upload processed embeddings. The edge server uses these embeddings to calculate node similarities and reconstruct a global topology graph. This "global flow" approach is more effective than local link prediction as it leverages shared embedding information to recover the true underlying structure.

Versatile Assessor with Negative Sampling

To ensure the quality of reconstructed links, an autoencoder-based assessor is implemented. It incorporates a negative sampling mechanism to filter out noise and irrelevant features. By identifying "negative" attributes, the model focuses on refined information that directly contributes to the classification task, preventing synthetic noise from degrading performance.

Overview of the SpreadFGL architecture, highlighting the interaction between edge servers and clients.

Distributed Multi-Edge Architecture

SpreadFGL replaces the centralized server with a distributed network of edge servers. These servers communicate via a ring topology, sharing regional knowledge and balancing the computational load. This decentralized approach not only improves scalability but also accelerates model convergence through a distributed loss function.

Graphic Patchers

On the client side, "graphic patchers" utilize the topology information provided by the edge servers to "repair" local subgraphs. This allows for proper feature propagation across reconstructed links during local GNN training.

Experimental Evaluation

The framework was evaluated on four benchmark datasets: Cora, Citeseer, WikiCS, and CoauthorCS. To validate real-world applicability, a hardware testbed was constructed using NVIDIA Jetson TX2s as edge servers and Raspberry Pi 4Bs as clients.

Key Results

SpreadFGL demonstrates superior performance across multiple dimensions:

  • Accuracy and F1-Score: SpreadFGL and its centralized variant (FedGL) outperformed the state-of-the-art FedSage+ by approximately 12.78% in accuracy and 14.71% in Macro F1-score.
  • Convergence Speed: The distributed architecture significantly reduces the number of iterations required to reach peak accuracy compared to standard FedAvg-based approaches.
  • Robustness: The model remains effective even with low labeled data ratios, a common scenario in semi-supervised learning tasks.
Convergence analysis showing node classification accuracy over iterations for Cora and Citeseer datasets.

Conclusion

SpreadFGL successfully addresses the "data island" and server bottleneck problems in Federated Graph Learning. By combining adaptive neighbor generation with a distributed edge-client architecture, it provides a robust and scalable solution for training GNNs on fragmented data. Future work will focus on optimizing communication overhead for low-power IoT environments and extending the framework to handle dynamic graphs.

Summary updated

BibTeX
@inproceedings{zhong2024spreadfgl,
  author = {Luying Zhong and Yueyang Pi and Zheyi Chen and Zhengxin Yu and Wang Miao and Xing Chen and Geyong Min},
  title = {{SpreadFGL: Edge-Client Collaborative Federated Graph Learning with Adaptive Neighbor Generation}},
  booktitle = {INFOCOM},
  year = {2024},
  doi = {10.1109/infocom52122.2024.10621368},
}
Figure