基于图熵理论的图数据增强研究  

Study on Graph Data Augmentation Based on Graph Entropy Theory

作　　者：富坤 崔静远 党兴 成晓 应世聪 李建伟 FU Kun;CUI Jingyuan;DANG Xing;CHENG Xiao;YING Shicong;LI Jianwei(School of Artificial Intelligence and Data Science,Hebei University of Technology,Tianjin 300401,China;Tianjin Institute of Aerospace Mechanical and Electrical Equipment,Tianjin 300462,China;Tianjin Key Laboratory of Aerospace Intelligent Equipment Technology,Tianjin 300462,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401 [2]天津航天机电设备研究所,天津300462 [3]天津市宇航智能装备技术企业重点实验室,天津300462

出　　处：《计算机科学》2025年第5期149-160,共12页Computer Science

基　　金：国家自然科学基金(62072154);天津市科技计划项目(22JCYBJC01740);河北省重大科技成果转化专项(22280803Z)。

摘　　要：图数据增强是一种通过变换和扩充图结构和节点特征来增加训练数据多样性、提高图神经网络性能的技术。为了应对图数据增强面临的难以综合信息完整性、特征平滑性、图多样性和局部依赖关系的挑战,缓解图神经网络的过平滑和过拟合问题,提高其性能,提出了一种基于物理热力学中的熵理论的图数据增强模型(Neighbor Replacement Based on Graph Entropy,NRGE)。首先,引入了一种新的图熵定义,用于度量数据流形平滑度;基于减少图熵损失的思想,提出了一种新的数据增强策略,用于生成更多合适的训练数据。然后,通过增强节点的采样邻居,以保证数据增强的一致性;采用随机替换节点的一阶邻居为二阶邻居的方式,增加了数据增强的多样性。最后,引入了邻居约束正则化方法,通过约束增强后的邻居之间的预测一致性来提高模型性能。消融实验结果表明,通过保持三角形图案的信息结构,NRGE模型能够有效降低图熵损失,从而改善学习效果。在Cora,Citeseer和Pubmed 3个公开数据集上进行了节点分类实验,相较于基准模型,NRGE模型在Cora数据集上提升了1.1%,在Citeseer数据集上提升了0.8%,在Pubmed数据集上略微降低了0.4%。结果表明,NRGE模型有效改善了图神经网络的性能,提高了其泛化能力。Graph data augmentation,as a technique aiming to enhance the performance of graph neural networks,involves transforming and expanding the graph structure and node features to increase the diversity and quantity of training data.The integrity of information structures,the smoothness of feature manifold,the diversity of graph,and local dependencies are difficult to comprehensively considered in graph data augmentation.Additionally,over-smoothing and over-fitting problems exist in the training of graph neural networks,which limit their learning capabilities.To address these issues,a graph data augmentation model(NRGE)based on the entropy theory in thermodynamics is proposed.Firstly,a novel definition of graph entropy is introduced to measure the smoothness of the feature manifold.A new data augmentation strategy,whose main idea is to reduce the loss of graph entropy is proposed to generate more appropriate training data.Secondly,the sampling neighbors of the nodes are augmented to ensure the consistency of data augmentation.To increase the diversity of data augmentation,the first-order neighbors of nodes are randomly replaced with their second-order neighbors.Finally,a neighbor-constrained regularization method is introduced,which improves model performance by enforcing prediction consistency between augmented neighbors.Ablation experiments show that the NRGE model effectively reduces the loss of graph entropy by preserving the information structure of triangles,thereby improving learning effect.Three real datasets are trained by the NRGE model.The obtained low-dimensional representation is applied to node classification.Compared with the baseline methods,the NRGE model achieves a performance improvement of 1.1%on the Cora dataset,0.8%on the Citeseer dataset,and a slight decrease of 0.4%on the Pubmed dataset.The experimental results show that the NRGE model can significantly enhance the performance of graph neural networks and improve the generalization ability.

关 键 词：图熵 图数据增强 邻居替换 一致性和多样性 结构增强 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]