ADPFedGNN: Adaptive Decoupling Personalized Federated Graph Neural Network

Zeli Guan, Yawen Li, Junping Du, Runqing Tang, Xiaolong MengIJCAI 2025 · DOI

Abstract

Federated Graph Neural Networks (FGNNs) enable collaborative learning on decentralized graph data while preserving privacy. However, existing FGNNs struggle with parameter coupling—where global and local knowledge interfere—and neighbor noise arising from non-IID data distributions. We propose ADPFedGNN, a novel framework that decouples global and local parameters through an automated masking mechanism and optimizes neighbor selection using reinforcement learning. By employing mutual information minimization, ADPFedGNN ensures the independence of extracted features. Our evaluation on standard citation benchmarks shows that ADPFedGNN significantly outperforms state-of-the-art methods, achieving up to a 12.45% improvement in recall.

Introduction

The rise of decentralized graph data has spurred interest in Federated Graph Neural Networks (FGNNs). Despite their promise, two critical issues hinder performance: parameter coupling and neighbor noise. Parameter coupling occurs when a single set of weights attempts to capture both general global patterns and specific local nuances, leading to mutual interference. Neighbor noise stems from the non-IID nature of graphs, where neighbors beneficial for global tasks may be detrimental to local personalization. ADPFedGNN addresses these by explicitly decoupling parameters and adaptively selecting neighbors.

Methodology

ADPFedGNN is composed of three primary modules:

  1. Reinforced Federated Adaptive Neighbor Selection: A Deep Q-Network (DQN) is trained to select neighbors dynamically. By incorporating a task-specific indicator in its state, the DQN learns to differentiate between neighbors that support global generalization and those that enhance local personalization.
  2. Federated Mask-based Parameter Decoupling: The framework utilizes two learnable masks ($M_{global}$ and $M_{local}$) to partition the GNN's parameter space. To prevent information leakage between the global and local branches, we minimize mutual information using the Contrastive Log-ratio Upper Bound (CLUB).
  3. Personalized Federated Masked Parameter Aggregation: Communication efficiency and personalization are maintained by only transmitting global-masked parameters to the server. Upon receiving the global update, clients apply an inverted mask to ensure their specialized local weights remain intact.
Overview of the ADPFedGNN framework architecture.

Experimental Results

We evaluated ADPFedGNN on Cora, CiteSeer, and PubMed datasets under various partitioning strategies (Louvain and Dirichlet). ADPFedGNN consistently outperformed eight baselines, including FedAvg, FedProx, and specialized FGNN methods like FedALA. Notably, it achieved average gains of 5.66% in Accuracy and 12.45% in Recall over the strongest baseline. The improvements were consistent across different GNN backbones, such as GraphSAGE and GAT.

Robustness to Heterogeneity

A key strength of ADPFedGNN is its resilience to data heterogeneity. As shown in the figure below, as the Dirichlet parameter $\alpha$ decreases (indicating higher imbalance), ADPFedGNN maintains a significantly more stable performance profile than competing methods.

Performance stability across varying data heterogeneity.

Conclusion

ADPFedGNN provides a robust solution for personalized federated graph learning by decoupling model parameters and intelligently filtering neighbor noise. While the reinforcement learning and mutual information components introduce some computational overhead, the performance gains and robustness to non-IID data make it a compelling choice for real-world federated graph applications. Future work will focus on optimizing the computational efficiency for large-scale industrial graphs and exploring hyperparameter sensitivity.

Summary updated

BibTeX
@inproceedings{guan2025adpfedgnn,
  author = {Zeli Guan and Yawen Li and Junping Du and Runqing Tang and Xiaolong Meng},
  title = {{ADPFedGNN: Adaptive Decoupling Personalized Federated Graph Neural Network}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/585},
}
Figure