Research On Semi-supervised Graph Node Classification Method Fused With Local Neighborhood Information

As a popular research direction in the field of graph analysis and mining,node classification has been widely used to analyze social networks and citation networks,etc.However,practical applications often face the problem of "less labeled data and more unlabeled data".Manually labeling data is time-consuming and labor-intensive,which is obviously not suitable for today's big data era.Therefore,it is of strong practical significance how to fully exploit and fuse key graph structure information to achieve the semi-supervised classification of graph nodes more effectively.Currently,deep learningbased network representation methods have achieved excellent results.However,in each update of the node vector representation,most existing works are based only on the information of directly connected neighboring nodes,ignoring the information of nondirectly connected local structures,which results in the absence of some useful information.Even though higher-order information can be involved by stacking multilayer networks,the presence of problems such as over-smoothing makes most methods limited by the depth of the model.In addition,most of the work usually utilizes only fixed-order neighbor information when integrating information of neighboring nodes.They cannot flexibly adapt to the range of neighbors according to the specific situation,so that the fused neighbor node information is not accurate enough.Therefore,to address the main problems and shortcomings at present,this paper proposes two improved graph convolutional network models incorporating abundant local neighborhood information for semi-supervised node classification tasks.The main work is summarized as follows:(1)A two-order neighborhood approximate spectral convolution network model based on Laplace polynomials is proposed.Combining the Chebyshev truncated expansion and the normalized Laplacian matrix,by deriving and simplifying the twoorder approximate spectral convolution module,an improved graph convolution formula that integrates abundant local structure information is given.When updating the node representation,the model can not only fuse the effective information of the direct neighborhood,but also gather the local structure information from the indirect neighborhood.Secondly,it demonstrates how to use the resulting model for the semisupervised classification task of graph nodes.A large number of quantitative and qualitative experiments are conducted on classic datasets.The results show that the proposed method in this paper achieves better classification results compared with the existing popular methods,verifying the effectiveness of the model.(2)A graph convolutional network model based on neighborhood adaptability is proposed.Combined with the energy diffusion theory,a neighborhood-adaptive convolution kernel is constructed on the graph using the energy heat kernel for better portraying the correlation between different nodes,which allows the model to determine the neighborhood range of nodes adaptively according to the given scale parameters.On this basis,by further introducing a threshold mechanism to filter part of the noisy node information in the adaptive neighborhood to reserve neighbor nodes that are more relevant to the target node,the model can more efficiently fuse the effective neighborhood node information in the convolution process.Extensive experiments are conducted on several classical citation datasets,and the results show that the proposed model achieves better classification results in multiple node classification tasks.Even in the case of low labeling rate,the model still shows some advantages,which illustrates the effectiveness and stability of the model.