Unsupervised multiplex graph diffusion networks with multi-level canonical correlation analysis for multiplex graph representation learning  

作　　者：Sichao FU Qinmu PENG Yange HE Baokun DU Bin ZOU Xiao-Yuan JING Xinge YOU 

机构地区：[1]School of Electronic Information and Communications,Huazhong University of Science and Technology,Wuhan 430074,China [2]Platform Operation and Marketing Center,JD Retail,Beijing 100176,China [3]Faculty of Mathematics and Statistics,Hubei Key Laboratory of Applied Mathematics,Hubei University,Wuhan 430062,China [4]School of Computer Science,Wuhan University,Wuhan 430072,China

出　　处：《Science China(Information Sciences)》2025年第3期62-75,共14页中国科学(信息科学)(英文版)

基　　金：supported in part by National Key Research and Development Program of China(Grant No.2022YFF0712300);National Natural Science Foundation of China(Grant No.62172177);Knowledge Innovation Program of Wuhan-Shuguang;Fundamental Research Funds for the Central Universities(HUST)(Grant No.2022JYCXJJ034);Open Research Fund from Shandong Provincial Key Laboratory of Computer Network(Grant No.SKLCN-2021-02)。

摘　　要：In recent years,unsupervised multiplex graph representation learning(UMGRL)has received increasing research interest,which aims to learn discriminative node features from the multiplex graphs supervised by data without the guidance of labels.Although these designed UMGRL methods have obtained great success in various graph-related tasks,most existing UMGRL models still have the following issues:highly depending on complex self-supervised strategies(i.e.,data augmentation,pretext tasks,and negative pairs sampling),restricted receptive fields,and only aggregating low-frequency information between nodes.In this paper,we propose a simple unsupervised multiplex graph diffusion network(UMGDN)with the aid of multi-level canonical correlation analysis to solve the above issues.Specifically,we first decouple the feature transform and propagation processes of the graph convolution layer to further improve the generalization of the learnable parameters.And then,we propose adaptive diffusion propagation to capture long-range dependency relationships between nodes,not the local neighborhood interactions.Finally,a multi-level canonical correlation analysis loss on both the feature transform and propagation processes is proposed to maximize the correlation of the same node features from multiple graphs for guiding model optimization.Compared to the existing UMGRL models,our proposed UMGDN does not need to introduce any data augmentation,negative pairs sampling techniques,complex pretext tasks,and also adaptively aggregates the optimal frequency information between nodes to generate more robust node embeddings.Extensive experiments on four popular datasets and two graph-related tasks demonstrate the effectiveness of the proposed method.

关 键 词：unsupervised multiplex graph representation learning graph neural networks node classification node clustering 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] O157.5[自动化与计算机技术—控制科学与工程]