CSC-GCN:Contrastive semantic calibration for graph convolution network

作     者：Xu Yang Kun Wei Cheng Deng 

作者机构：School of Electronic EngineeringXidian UniversityXi'an 710071China 

出 版 物：《Journal of Information and Intelligence》 (信息与智能学报（英文）)

年 卷 期：2023年第1卷第4期

页      面：295-307页

学科分类：08[工学] 081104[工学-模式识别与智能系统] 0811[工学-控制科学与工程] 

基　　金：supported by Joint Fund of Ministry of Education of China(8091B022149) Key Research and Development Program of Shaanxi(2021ZDLGY01-03) National Natural Science Foundation of China(62132016,62171343,62071361 and 62201436) Fundamental Research Funds for the Central Universities(ZDRC2102 and ZYTS23135). 

主　　题：Graph convolution network Contrastive learning Semi-supervised learning 

摘      要：Graph convolutional networks(GCNs)have been successfully applied to node representation learning in various real-world applications.However,the performance of GCNs drops rapidly when the labeled data are severely scarce,and the node features are prone to being indistinguishable with stacking more layers,causing over-fitting and over-smoothing problems.In this paper,we propose a simple yet effective contrastive semantic calibration for graph convolution network(CSC-GCN),which integrates stochastic identity aggregation and semantic calibration to overcome these weaknesses.The basic idea is the node features obtained from different aggregation operations should be similar.Toward that end,identity aggregation is utilized to extract semantic features from labeled nodes,while stochastic label noise is adopted to alleviate the over-fitting problem.Then,contrastive learning is employed to improve the discriminative ability of the node features,and the features from different aggregation operations are calibrated according to the class center similarity.In this way,the similarity between unlabeled features and labeled ones from the same class is enhanced while effectively reducing the over-smoothing problem.Experimental results on eight popular datasets show that the proposed CSC-GCN outperforms state-ofthe-art methods on various classification tasks.