基于随机重构图结构的图神经网络分类算法

The classification algorithm of a graph neural network based on a random reconstructed graph structure

目的:为了解决在特征聚合过程中的确定性传播所导致的节点相似性破坏和节点对邻域依赖性高的问题,构建基于随机重构图结构的图神经网络分类算法。方法:首先,随机特征变换根据学习的权重值对随机保留的部分节点特征进行增强,生成随机特征。然后,利用生成的特征计算融合系数,对原始图和k近邻图进行自适应融合,重构出随机图结构。最后,将提取的多支浅层特征加入到重构图结构的卷积层,使模型随着层数的加深可得到浅层信息的补充。此外,对联合优化分类损失和自监督学习损失,保持节点相似性和平滑性。结果:与其他节点分类方法在Cora、Citeseer和Pubmed数据集上进行半监督实验和全监督实验结果对比,本文的算法精度提高了0.9%~2.3%。结论:基于随机重构图结构的网络分类算法在节点分类任务中取得较好的性能。

Aims:In order to solve the problems of deterministic propagation in the process of feature aggregation,the damage of node similarity and the high dependence of nodes on the neighborhood,a classification algorithm of the graph neural network based on a random reconstructed graph structure was constructed.Methods:First,random features were generated by random feature transformation,which enhanced some of the reserved node features according to the learned weight values.Then,the fusion coefficient was calculated according to the generated features;and the original graph and the k-nearest neighbor graph were adaptively fused to reconstruct the random graph structure.Finally,the shallow features extracted from multiple branches were added to the convolution layer of the reconstructed graph structure,so that the model could be supplemented with shallow information as the number of layers deepened.Furthermore,the classification loss and self-supervised learning loss were jointly optimized to preserve node similarity and smoothness.Results:Compared with other node classification methods on the Cora,Citeseer and Pubmed datasets,the accuracy improved by 0.9%~1.3%.Conclusions:The network classification algorithm based on the random reconstructed graph structure can achieve good performance in the node classification task.