实验技术与管理2025，Vol.42Issue(11)：91-100,10.DOI:10.16791/j.cnki.sjg.2025.11.010
基于图扩散卷积和伪标签的抗噪声图神经网络研究
Research on anti-noise graph neural network based on graph diffusion convolution and pseudo-labels
摘要
Abstract
[Objective]Traditional graph neural networks(GNNs)aggregate information from their first-order neighbor nodes through a local neighborhood aggregation mechanism to update the nodes'own representations;while this mechanism offers efficient computational capability,it also causes noisy label information to spread rapidly across the graph structure,contaminating the representations of unlabeled nodes,and in semi-supervised learning scenarios,the limited labeled data further exacerbates the impact of noise,making the model prone to overfitting,with existing anti-noise methods often failing to fully utilize the global information of the graph or struggling to maintain stable performance under high noise levels,thus this study aims to develop a more robust anti-noise method to solve the problem that unlabeled nodes are adversely affected by noisy labels,which significantly degrades the performance of GNNs.[Methods]The proposed method constructs an anti-noise graph neural network based on graph diffusion convolution(GDC)and pseudo-labels,where the GDC mechanism(adopting Personalized PageRank and Heat Kernel as diffusion methods)breaks the limitation of local neighborhood aggregation to capture global graph information,and the framework integrates three core components—an edge predictor,a pseudo-label miner,and a GNN classifier—to first transform the original graph into a new graph through sparsification and graph diffusion,then connect feature-similar unlabeled and labeled nodes via the edge predictor to form an intermediate graph,train the pseudo-label miner on this intermediate graph to identify high-confidence pseudo-labeled nodes that are combined with original labeled nodes to form an extended labeled node set,and finally use the GNN classifier to fuse node attributes and graph topology for accurate node classification,with a comprehensive loss function balancing the reconstruction loss of the edge predictor,the classification loss of the pseudo-label miner,and the classification loss of the GNN classifier.[Results]Experimental results show that compared with mainstream anti-noise models(such as GCN,GAT,and self-training),the proposed method achieves significant performance improvements in semi-supervised node classification tasks,with a 5%to 10%improvement even under high noise levels;the model exhibits stronger anti-interference performance against Uniform noise than Pair noise,and demonstrates powerful generalization ability under various noise conditions regardless of the dataset type and size,including academic citation network datasets(Cora,CiteSeer),large-scale complex academic literature network datasets(DBLP),medical literature datasets(PubMed),and e-commerce product association network datasets(Amazon-Photo);additionally,the model shows good stability with relatively small fluctuations in accuracy during experiments,indicating it is less affected by random factors.[Conclusions]The proposed anti-noise graph neural network based on graph diffusion convolution and pseudo-labels can stably perform,effectively address the problem of noisy label propagation in GNNs,and improve the model's robustness in semi-supervised learning scenarios;the combination of graph diffusion convolution and pseudo-labels is effective for node classification tasks on graph-structured data with different noise types and levels,and has good generalization performance,which is superior to other models under the same conditions,proving the advancement of the method;since the model's performance is affected by diffusion methods,hyperparameters,noise types,and noise levels,it is crucial to select appropriate hyperparameters according to different scenarios to ensure optimal performance.关键词
图神经网络/伪标签/图扩散卷积/噪声标签Key words
graph neural network/pseudo labels/graph diffusion convolution/noisy labels分类
信息技术与安全科学引用本文复制引用
李崇,刘奇磊,韩毅,孙高飞,罗盟千,雷文利..基于图扩散卷积和伪标签的抗噪声图神经网络研究[J].实验技术与管理,2025,42(11):91-100,10.基金项目
辽宁省高校基本科研项目(LJ212410149007) （LJ212410149007）
辽宁省自然科学基金项目(2025-MS-143) （2025-MS-143）
河南省科技攻关计划(242102220038) （242102220038）
河南省自然科学基金项目(242300420459) （242300420459）
河南省高校重点科研项目(23A520048) （23A520048）
