FedBG: Proactively Mitigating Bias in Cross-Domain Graph Federated Learning Using Background Data

Sheng Huang, Lele Fu, Tianchi Liao, Bowen Deng, Chuanfu Zhang, Chuan ChenIJCAI 2025 · DOI

Abstract

Federated Graph Learning (FGL) faces significant challenges due to domain shift, where decentralized clients possess graphs with varying node features and topological structures. Existing reactive calibration methods often fail to preserve knowledge integrity during model aggregation. We present FedBG (Federated graph learning with BackGround data), a framework that proactively mitigates domain bias. By sharing privacy-preserving, condensed "background graphs," clients can align their models with a global context prior to local training. Evaluated on real-world benchmarks like Twitch and Facebook100, FedBG demonstrates superior accuracy, stability, and communication efficiency compared to state-of-the-art FGL methods.

Introduction

Federated Graph Learning (FGL) aims to train Graph Neural Networks (GNNs) across decentralized clients without exposing raw data. However, FGL is uniquely susceptible to domain shift. Unlike standard image or text data, graphs couple node features with complex topological structures. When clients draw data from different sources—such as social networks in different regions—they exhibit both "feature skew" and "topology skew."

Most existing FGL methods attempt to address this bias reactively, adjusting the global model after it has been influenced by biased local updates. FedBG shifts this paradigm by introducing proactive calibration, ensuring local models are aligned with the global data distribution before local training begins.

Methodology

FedBG introduces the concept of Background Data to facilitate proactive calibration. The framework operates through four primary stages:

  1. Graph Condensation: Clients use gradient matching to compress large private graphs into small "background graphs." These condensed versions retain essential feature and structural information while protecting raw data privacy.
  2. Privacy-Preserving Mixing: Condensed graphs are uploaded to a central server, which clusters and mixes them. This process "blurs" individual client contributions while creating a shared reference representing the collective network knowledge.
  3. Diversity Restoration: To prevent the mixed data from becoming too homogeneous, FedBG uses an adversarial graph generation method. This restores the unique characteristics of different domains, ensuring the model adapts to diverse environments.
  4. Proactive Local Tuning: Before standard local training, clients use the distributed mixed background graphs to "pre-tune" their models. This aligns the local model with the global distribution, preventing significant drift during the subsequent local training phase.
The FedBG framework architecture, illustrating the flow from local graph condensation to global background data mixing and proactive tuning.

Experimental Setup and Datasets

The authors evaluated FedBG against traditional Federated Learning (FL) and specialized FGL baselines using two major real-world benchmarks:

  • Twitch: A social network dataset covering six regional domains (EN, ES, FR, PT, RU, and DE).
  • Facebook100: A dataset of university social networks. A "lite" version (sampling 10% of nodes) was used to test robustness in data-scarce scenarios.

Key Results

FedBG consistently outperformed state-of-the-art methods across all benchmarks:

  • Accuracy Gains: On the Twitch dataset, FedBG achieved 65.69% accuracy, a 5.14% improvement over the standard FedAvg baseline. On Facebook100-lite, it reached 57.19%, outperforming the specialized FedGTA method by nearly 2%.
  • Stability and Convergence: Convergence curves indicate that FedBG reaches higher accuracy faster and maintains more stable performance across communication rounds compared to reactive methods like FedProx or MOON.
  • Communication Efficiency: FedBG maintains high accuracy even when communication frequency is reduced by up to 20x. This is because the background data provides a persistent environmental context that keeps local models "on track" between rounds.
Convergence curves showing classification accuracy over communication rounds for Twitch and Facebook100-lite datasets.

Discussion and Future Work

FedBG demonstrates that data-centric calibration is more effective than model-centric approaches for FGL. By treating topology as a first-class citizen and calibrating proactively, the framework preserves knowledge integrity across heterogeneous domains.

Future research may explore the balance between privacy-preserving mixing and data diversity, as well as extending the background data approach to other graph tasks such as link prediction or graph-level classification.

Summary updated

BibTeX
@inproceedings{huang2025fedbg,
  author = {Sheng Huang and Lele Fu and Tianchi Liao and Bowen Deng and Chuanfu Zhang and Chuan Chen},
  title = {{FedBG: Proactively Mitigating Bias in Cross-Domain Graph Federated Learning Using Background Data}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/602},
}
Figure