智能系统学报2025，Vol.20Issue(5)：1243-1255,13.DOI:10.11992/tis.202409034
多双曲空间传递图解耦表示学习
Graph disentanglement representation learning based on propagation in multiple hyperbolic spaces
摘要
Abstract
There are two salient issues of existing graph representation learning methods.First,there is a dearth of fine-grained neighborhood modeling,which neglects the multifaceted semantic entanglements in the neighborhood struc-tures.Second,the spatial metric employed in graph representation learning presents a significant challenge,since Euc-lidean space may not constitute the optimal framework for quantifying node representations.To solve these challenges,this study proposes a novel representation propagation and prediction mechanism within multiple hyperbolic spaces,thereby achieving disentangled graph representation learning under multifaceted hyperbolic spatial metrics.Within the proposed framework,the original topological structure is iteratively refined through node representations,yielding propagation matrices embedded in a hyperbolic space.Furthermore,based on a mixture-of-experts design,hyperbolic la-bel propagation networks at different resolutions are treated as expert networks,enabling the discovery of node connec-tion patterns induced by different latent factors.Experimental results on multiple real-world datasets show that the pro-posed method achieves classification accuracies of 32.3％and 59.5％on the Squirrel and Crocodile datasets,respect-ively.Additionally,visualization experiments further demonstrate the effectiveness of the proposed approach.关键词
图表示学习/图解耦/双曲空间/图神经网络/标签传递/混合专家系统/拓扑细化/多分辨率Key words
graph representation learning/graph disentanglement/hyperbolic space/graph neural networks/label propagation/mixture of experts/topology refinement/multiresolution分类
信息技术与安全科学引用本文复制引用
郑帅,彭奏章,朱振峰,赵耀..多双曲空间传递图解耦表示学习[J].智能系统学报,2025,20(5):1243-1255,13.基金项目
中央高校基本科研业务费项目(2024XKRC088) （2024XKRC088）
国家自然科学基金项目(62476022) （62476022）
北京市自然科学基金青年科学基金项目(4254085). （4254085）
