附加特征图增强的图卷积神经网络

Graph Convolutional Neural Networks with Additional Feature Graph

近年来,图卷积网络(Graph Convolutional Network,GCN)凭借其简单的网络结构、在图上任务中展现出的优异性能,受到了学术界和工业界的广泛关注.然而GCN也存在着在浅层时信息传播范围过小、特征提取不充分的缺陷.针对这一问题,本研究提出附加特征图模型(Additional Feature Graph,AFG).AFG通过引入图的节点结构特征(度特征),对度相同的节点随机增加连边、缩短信息传播距离.AFG并不是独立的图神经网络模型,而是作为一种附加技术与GCN及其相关模型配合使用.实验显示,在Cora、Citeseer、Pubmed数据集上AFG能够对浅层主干模型实现显著性能增益,帮助主干模型性能超越了其他以提升模型特征提取能力、改善欠传播情况为目的进行设计的模型.本研究进一步分析了AFG与DropEdge——一种随机切断原始图连边的附加技术——的区别与联系,并通过实验证明附加特征图模型与DropEdge模型共同使用的可行性,以及两者间存在一定的互补性.结合使用两种附加技术可以实现更大的节点分类准确度增益.

Graph structures are suitable for the modeling of complex interactions and relations.Therefore,graphs are widely used in data representation such as molecules,chemical compounds,citation networks,social networks,traffic web,and knowledge graphs.In light of the great success of neural networks in image understanding and natural language processing,there has been a rising interest in Graph Neural Networks(GNN)for the study of learning on graphs.Among the popular GNNs,Graph Convolutional Networks(GCN),highlighted for their simple network structures and excellent performance with graphs,have attracted wide attention and become a promising direction.However,the limitation on message passing distance deteriorates the performance of GCN.To address this problem,people propose constructing deep GCN models to improve propagation.However,as the depth of GCN increases,node features become oversmoothed or over-squashed.This leads to a sharp decrease in the model performance.Though different deep GCN models have been proposed to tackle the over-smoothing or over-squashing problem,the inherent problems of the message passing mechanism are less explored.The problems of message passing in GCN include:1)having unreliable paths in the original input graphs which pass information with low signal-to-noise ratios,2)reaching limited message passing extent which gives rise to the less expressive feature representation,and 3)lacking explicit learning fashion on structural features.To this end,some people seek to directly improve the message passing in shallow GCN concerning the limitation in robustness,message passing extent,or feature diversity.But all these studies fail to tackle three problems in a single model.In this paper,we propose a novel model named AFG(Additional Feature Graph),to improve the message passing over robustness,message passing extent,and feature diversity.Specifically,AFG can inject the structural features of the input graph into the message passing process and randomly add edges between node pairs that have the same degree.The degree feature represents the first-order topological structure of individual nodes.Connecting nodes with the same degree allow explicit learning on structural features.As a lightweight and general technique,our AFG model can be easily plugged into GCN and its related models,bringing extra improvements.Experimental results on three datasets demonstrate the efficiency of AFG.AFG-aided models outperform shallow backbones and related GCN-based models on Cora,Citeseer,and Pubmed.AFG achieves 0.69%averaged improvement on 2-layer models,0.57%on 4-layer models,1.18%on 6-layer models,and 0.99%on 8-layer models.The improvements are proven to be significant with hypothesis testing.We also provide detailed experimental analysis on AFG with both synthetic datasets and real world datasets.The corresponding experimental results demonstrate that AFG can improve the connectivity and shorten the average path length of the input graphs.With AFG,nodes can get more informative features from their neighbors.Compared to GCN,AFG reaches a broader message passing extent in shallow model structures.Furthermore,we provide experiments verifying that AFG and DropEdge(another plug-and-play technique for GCN models)are complementary to each other and can be combined to achieve better performance.