FedHERO: A Federated Learning Approach for Node Classification Task on Heterophilic Graphs

Zihan Chen [email protected], Xingbo Fu [email protected], Yushun Dong [email protected], Jundong Li [email protected], Cong Shen [email protected] · 2025

Read FedHERO: A Federated Learning Approach for Node Classification Task on Heterophilic Graphs by Zihan Chen [email protected], Xingbo Fu [email protected], Yushun Dong [email protected], Jundong Li [email protected], Cong Shen [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-21206.txt.

Introduction

Zihan Chen, Xingbo Fu, Yushun Dong, Jundong Li, and Cong Shen present FedHERO, a federated method for labeling nodes in graphs whose neighbors tend to differ. Their paper appeared in two thousand twenty-five. Federated graph learning trains shared models across private data holders. Most methods assume neighbors resemble each other, as friends often share tastes.

Real graphs often break that rule: fraudsters link to honest users, and unlike nodes connect. When clients hold graphs with different mixing patterns, their local lessons clash. Merging them can wreck the global model. FedHERO's two-channel network learns the structure of each local graph as it goes. It separates the patterns that travel across graphs from those that stay local. Client models gain from both their own and the shared structural insight. The authors claim a new standard for heterophilic federated learning.

Catalog record

CREATORZihan Chen [email protected], Xingbo Fu [email protected], Yushun Dong [email protected], Jundong Li [email protected], Cong Shen [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-21206

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

FedHERO: A Federated Learning Approach for Node Classification Task on Heterophilic Graphs Zihan Chen [email protected], Xingbo Fu [email protected], Yushun Dong [email protected], Jundong Li [email protected], Cong Shen [email protected] Abstract Graph neural networks (GNNs) have shown significant success in modeling graph data, and Federated Graph Learning (FGL) empowers clients to collaboratively train GNNs in a distributed manner while preserving data privacy. However, FGL faces unique challenges when the general neighbor distribution pattern of nodes varies significantly across clients. Specifically, FGL methods usually require that the graph data owned by all clients is homophilic to ensure similar neighbor distribution patterns of nodes. Such an assumption ensures that the learned knowledge is consistent across the local models from all clients. Therefore, these local models can be properly aggregated as a global model without undermining the overall performance. Nevertheless, when the neighbor distribution patterns of nodes vary across different clients (e.g., when clients hold graphs with different levels of heterophily), their local models may gain different and even conflict knowledge from their node-level predictive tasks. Consequently, aggregating these local models usually leads to catastrophic performance deterioration on the global model. To address this challenge, we propose FedHERO , an FGL framework designed to harness and share insights from heterophilic graphs effectively. At the heart of FedHERO is a dual-channel GNN equipped with a structure learner, engineered to discern the structural knowledge encoded in the local graphs. With this specialized component, FedHERO enables the local model for each client to identify and learn patterns that are universally applicable across graphs with different patterns of node neighbor distributions. FedHERO not only enhances the performance of individual client models by leveraging both local and shared structural insights but also sets a new precedent in this field to effectively handle graph data with various node neighbor distribution patterns. We conduct extensive experiments to validate the superior performance of FedHERO against existing alternatives. 1 Introduction Graph Neural Network (GNN) aims to extract informative patterns from graphs ( Xia et al., 2021 ; Wu et al., 2020b ; Zhou et al., 2020 ; Scarselli et al., 2008 ; Xu et al., 2018a ) . However, graph data may contain sensitive information about the involved individuals (e.g., a user’s race in a social network ( Dwork et al., 2012 ; Beutel et al., 2017 ) ). As a consequence, sharing graph data across different stakeholders is raising concerns due to the potential risk of privacy leakage, which further prevents training GNNs in a centralized manner ( Cong & Mahdavi, 2022 ; Dai et al., 2022 ; Mei et al., 2019 ; Zhu & Zhu, 2020 ) . To properly handle the above-mentioned problems, Federated Graph Learning (FGL) has emerged as a popular collaborative learning framework, which enables participants (clients) to derive insights from distributed graph sources while upholding stringent privacy safeguards ( Liu et al., 2022a ; Lou et al., 2021 ; Zhu et al., 2022 ; Zhang et al., 2022 ; Peng et al., 2021 ; Chen et al., 2021b ) . A typical scenario of FGL is that each client possesses a single graph (local graph) and seeks to collaboratively train a global GNN model for certain tasks (e.g., node classification tasks ( Xiao et al., 2022 ; Bhagat et al., 2011 ) ) without sharing the raw graph data. In such scenarios, current FGL approaches implicitly rely on the assumption of homophilic local graphs ( Zhu et al., 2020 ; Ma et al., 2021 ) , i.e., nodes within a client’s graph are more likely to form connections with other nodes of the same class ( Ma et al., 2019 ; Tang et al., 2009b ) . Specifically, when clients hold homophilic graphs, most of the neighbors of each node will have the same class as this center node ( Ying et al., 2018 ) . The GNNs trained on different clients will thus capture a similar tendency, and aggregating these GNNs usually makes the global model in FGL better learn such a tendency and benefit the generalization ability of the global model ( Baek et al., 2023 ; Zhang et al., 2021 ) . Figure 1 : An example of financial transaction networks in two banks. The edges in the figure represent transaction records. Due to the diverse customer financial habits in distinct networks, customers (nodes) within the same risk classification (label) across different banks have diverse transactional relationship patterns (neighbor distribution). Nevertheless, a critical issue regarding the above-mentioned assumption is that it overlooks the prevalence of heterophilic graphs ( Pandit et al., 2007 ; Zheng et al., 2022 ; Zheng et al., 2023 ; Bo et al., 2021 ; Yang et al., 2021 ; Chien et al., 2020 ; Liu et al., 2023a ) , where nodes belonging to different classes tend to be connected. For instance, in financial transaction networks, nodes denote bank clients, where node labels are their credit risk and edges represent the financial interactions between them. Such a network can be heterophilic, and we show an example in Figure 1 . Specifically, a ‘Moderate Risk’ customer in Bank A may transact with multiple ‘Low Risk’ individuals, whereas in Bank B, the ’Moderate Risk’ customer engages primarily with ’High Risk’ counterparts. When training GNNs to predict customer risk, Bank A’s GNN learns the connection between ’Moderate Risk’ and ’Low Risk,’ while Bank B’s GNN captures the connection between ’Moderate Risk’ and ’High Risk.’ Directly aggregation models could lead to poor performance in predicting ’Moderate Risk’ customers, as they fail to establish a consistent pattern that generalizes well across both banks. Consequently, this mismatch can result in performance degradation for FGL algorithms. To properly handle the challenge discussed above, in this paper, we introduce FedHERO , a novel framework for Fed erated Learning on He te ro philic Graphs. FedHERO is equipped with a carefully designed dual-channel GNN model, including the global channel and the local channel. The dual-channel design aims to aggregate GNNs that have learned similar patterns to facilitate knowledge sharing. To ensure each client obtains a GNN model aligned with others, we introduce a shared structure learning model that generates latent graphs across clients based on common patterns. By learning from and sharing the knowledge from these latent graphs, GNNs in the global channel capture similar tendencies, enhancing their overall performance. Meanwhile, the local channel operates directly on the original graph. With such a design, FedHERO facilitates the sharing of common knowledge across clients while leveraging the unique structural information in each client’s local graph to enhance overall model performance. This approach offers two key advantages: (1) The global channel, trained on the latent graph, reduces reliance on the original neighbor distribution. The GNN models trained on different clients could capture similar tendencies from the shared structure learning model, and aggregating these GNNs could exploit FGL’s potential. (2) Retaining the locally trained GNN on the original graph preserves sensitive information, enhancing privacy protection. We summarize the contributions of our paper below. • Problem Formulation. To the best knowledge, we take an initial step to formulate and handle the challenge arising from the implicit heterophilic assumption of FGL. • Algorithmic Design. We introduce an innovative FGL framework, named FedHERO , which empowers clients

Read the complete text