基于标签传播增强的多通道图卷积网络

Multi-channel Graph Convolutional Networks Enhanced by Label Propagation Algorithm

多数图卷积网络(GCN)模型通过设计高效的信息传递和保留方式提升节点分类任务的实验表现,忽略了节点标签信息在拓扑空间和属性空间的传播。针对上述问题,提出了一种基于标签传播算法(LPA)增强的多通道图卷积模型MGCN-LPA,同时增大同类节点在属性和拓扑空间的关系权重,改善节点间特征和标签信息的传播。首先,计算不同节点的属性相似度值,并采用k近邻算法生成属性关系图;然后,利用结合了GCN和LPA的图卷积层GCN-LPA提取属性图和属性关系图的潜在特征,生成拓扑节点表示和属性节点表示;最后,将拓扑和属性表示进行融合,并将生成的最终表示用于节点分类任务。在3个基准图数据集上进行实验,MGCN-LPA的实验表现能够匹配当前较为先进的基线模型,其在Cora和Citeseer数据集上的分类结果相比表现最优的基线模型提升了9.3%和12%。上述实验结果表明,MGCN-LPA能够增大同类节点间路径的权重,从而增强同类节点间的信息传递,提升节点分类任务的实验表现。此外,消融实验结果表明,与仅使用拓扑空间或者属性空间信息的变体相比,融合两类信息的MGCN-LPA能够充分提取和保留原始图中蕴含的潜在特征,提升模型的表征能力和泛化性。

Most graph convolutional networks(GCN)improve the experimental performance of node classification tasks by designing efficient methods for information propagation and preservation,while ignoring the propagation of node label information in the topological and attribute spaces.Aiming at the above problems,the paper proposes a multi-channel graph convolution model MGCN-LPA enhanced by the label propagation algorithm(LPA).The model enhances the propagation of node features and label information by increasing the weights of relationship between nodes of the same class in the attribute space and topology space.Firstly,it calculates the similarity values of different node attributes and generates an attribute relation graph using the k-nearest neighbor algorithm.Then,it combines the GCN and LPA in the graph convolution layer GCN-LPA to extract potential features from the attribute graph and attribute relation graph,generating topological node representations and attribute node representations.Finally,the method combines the topological and attribute representations and utilizes the final representation for node classification tasks.On three benchmark graph datasets,the experimental performance of MGCN-LPA can match the current state-of-the-art baseline models.The classification results on the Cora and Citeseer datasets show improvements of 9.3%and 12%respectively compared to the best-performing baseline.The experimental results demonstrate that MGCN-LPA can increase the weights of paths between nodes of the same class and enhance the propagation of information among nodes of the same class,thereby enhancing the performance of node classification tasks.In addition,the ablation experiments demonstrate that the fusion of both topological space and attribute space information in MGCN-LPA enhances the model’s representational capacity and ge-neralization compared to variants using only one type of information.This fusion allows for the full extraction and preservation of latent features present in the original graph.