Structure-Preserving Mixture-of-Experts: A graph sparse computation framework for node classification

2026-07-30

PRE

Ming Li

Yimeng Wang

Huan Xu

Bo Peng

车翔玖 (Xiangjiu Che) *

DOI：10.1016/j.knosys.2026.116693

Original

Original request for help

Save

Abstract

En 中文

Node classification is a core benchmark task in graph neural networks (GNNs). It aims to predict the class of each node in a graph and offers a diverse array of applications of graph-structured data. Some studies propose GNN frameworks that integrate the Mixture-of-Experts (MoE) approach for node classification to reduce computation and tackle issues in large-scale graphs, such as oversmoothing, redundant computation, and structural fragmentation. However, most existing methods rely on isolated node-choice routing. In these approaches, each node is routed independently, making expert computation domains poorly aligned with graph topology, which fragments the graph structure and limits structure-aware specialization of experts. Therefore, we propose Structure-Preserving MoE (SP-MoE), a framework that correlates expert activation with graph topology. Each expert selects the top- k core nodes via a structure-aware gate module and conducts message passing on local domains expanded from these core nodes. Meanwhile, a global expert module is designed to ensure the coverage of marginal nodes and the completeness of representations. Extensive experiments demonstrate that the proposed SP-MoE consistently achieves strong and competitive performance compared to classical GNNs and recent scalable graph architectures.

Keywords:

Graph neural network

Mixture-of-Experts

Node classification

Structure-Preserving routing

Global expert

Structure-Preserving Mixture-of-Experts: A graph sparse computation framework for node classification

Abstract

Journal

Organization

Cited Papers

Citing Papers