[论文解读] Heterogeneity-Aware Knowledge Sharing for Graph Federated Learning

Wentao Yu, Sheng Wan+3 more|arXiv (Cornell University)|Jan 29, 2026

Advanced Graph Neural Networks被引用 0

一句话总结

FedSSA 通过语义对齐和结构对齐，在图分布式学习中同时解决节点特征异质性和结构异质性，采用节点特征的变分聚类和基于谱能量的结构聚类，具有理论收敛性和显著的实验提升。

ABSTRACT

Graph Federated Learning (GFL) enables distributed graph representation learning while protecting the privacy of graph data. However, GFL suffers from heterogeneity arising from diverse node features and structural topologies across multiple clients. To address both types of heterogeneity, we propose a novel graph Federated learning method via Semantic and Structural Alignment (FedSSA), which shares the knowledge of both node features and structural topologies. For node feature heterogeneity, we propose a novel variational model to infer class-wise node distributions, so that we can cluster clients based on inferred distributions and construct cluster-level representative distributions. We then minimize the divergence between local and cluster-level distributions to facilitate semantic knowledge sharing. For structural heterogeneity, we employ spectral Graph Neural Networks (GNNs) and propose a spectral energy measure to characterize structural information, so that we can cluster clients based on spectral energy and build cluster-level spectral GNNs. We then align the spectral characteristics of local spectral GNNs with those of cluster-level spectral GNNs to enable structural knowledge sharing. Experiments on six homophilic and five heterophilic graph datasets under both non-overlapping and overlapping partitioning settings demonstrate that FedSSA consistently outperforms eleven state-of-the-art methods.

研究动机与目标

• Motivate and address the dual heterogeneity in Graph Federated Learning: node feature heterogeneity and structural heterogeneity.
• Propose FedSSA to share semantic and structural knowledge across clients while preserving privacy.
• Develop clustering and alignment mechanisms for both feature distributions and spectral graph structures.
• Provide theoretical convergence guarantees and empirical evidence on diverse datasets (homophilic and heterophilic).

提出的方法

• Introduce a variational model to infer class-wise node feature distributions per client and cluster clients by these distributions.
• Construct cluster-level representative distributions by moment matching (mean and covariance) and minimize KL divergence between local and cluster distributions to share semantic knowledge.
• Use spectral GNNs and a novel spectral energy measure to characterize each client’s structure; cluster clients on spectral energy using Grassmann manifold geometry.
• Build cluster-level spectral GNNs and align local spectral GNNs to cluster-level GNNs by matching spectral coefficients and enforcing regularization.
• Prove linear convergence of FedSSA under standard assumptions and define the error floor combining semantic and structural misalignment terms.

实验结果

研究问题

• RQ1 How can node feature heterogeneity be explicitly addressed in graph federated learning through semantic alignment?
• RQ2 How can structural heterogeneity be captured and mitigated via spectral information and alignment across clients?
• RQ3 What are the theoretical convergence properties of a dual-knowledge-sharing framework addressing both feature and structure heterogeneity?
• RQ4 Do semantic and structural alignments yield consistent gains across homophilic and heterophilic graphs under varying partitioning schemes?

主要发现

• FedSSA consistently outperforms eleven state-of-the-art methods across six homophilic and five heterophilic datasets.
• In heterophilic settings, FedSSA surpasses the second-best method (FedIIH) by 2.82 percentage points in classification accuracy.
• The framework achieves strong empirical performance under both non-overlapping and overlapping partitioning schemes across 66 scenarios.
• The authors provide a linear convergence guarantee to an O(E) neighborhood of the optimum, where E aggregates semantic and structural error terms.
• Extensive experiments on diverse datasets demonstrate stability with small standard deviations.
• The convergence and performance gains are grounded in explicit separation and alignment of semantic and structural knowledge.