Research And Application Of Homogeneous Graph Node Classification Method Based On Graph Convolution Neural Network

Graph-structured data is a data structure composed of nodes and edges,which can directly show the structure and relationship between complex data in real life.Therefore,how to efficiently analyze and apply graph structured data is an important research topic in the field of graph learning,and the classification of homogeneous graph nodes has become one of the hot issues in graph structured data research.In recent years,many researchers have proposed various graph convolutional neural network methods to solve the above homogeneous graph node classification problem.However,there are still some shortcomings to be solved in such methods.For example,when dealing with a complex homogeneous graph,the existing method cannot effectively extract complex information between nodes,and is difficult to meet the requirements of real-time performance and efficiency,resulting in a decrease in model accuracy.In addition,the existing method is weak in robustness,poor in generalization ability,sensitive to noise in homogeneous images,and easy to be disturbed to cause deviation of results.In order to solve the above problems,this thesis proposes several homogeneous graph node classification methods based on graph convolutional neural network to improve the shortcomings of the existing methods,and selects the enterprise safety production as a typical scenario,and verifies the effectiveness of the proposed method through the safety risk assessment task.Specific research contents and innovations are as follows:(1)In order to solve the problem of accuracy drop in current node classification methods when dealing with complex homogeneous graphs,this thesis proposes a node classification method based on feature adaptive graph convolutional neural network.In the process of feature transfer,an adaptive feature extraction strategy is designed,and the similar or different feature between nodes are extracted adaptively through positive and negative weights,so that the over-smoothing phenomenon is avoid.In addition,a hybrid attention mechanism is construct,which combines that advantage of the self-attention mechanism and the dot-product attention mechanism,so that the proposed model has faster operation speed and more effective parameter space.The semi-supervised node classification experiments on six representative public datasets Cora,Citseeker,Pubmed,Chameleon,Squirrel and Actor show that the accuracy of the proposed method is 84.6%、73.2%、80.0%、76.5%、67.3% and 37.2%,which have better expressive power than other SOTA baseline models,thus verifying the effectiveness of the method.(2)In order to solve the problems of poor robustness and poor generalization ability of existing node classification methods,a node classification method based on component-level attention graph convolution is proposed.According to this method,by constructing a component-level attention mechanism,neighbor node weight learning and feature extraction processes are separated,a feature similarity component learns the importance of each feature dimension,and a node similarity component learns the importance of neighbor nodes to a target node,so that node feature information is extracted to the maximum extent.In addition,dynamic graph connection strategy and attention module based on feature normalization are designed to expand the range of node feature extraction and reduce information loss.The semi-supervised node classification experiments on six datasets show that the proposed method outperforms other SOTA baseline models in accuracy rate of 84.9%、73.5%、80.2%、76.9%、70.7% and 37.3%,which verify that the proposed method can further optimize the model performance.(3)In order to verify the effectiveness of the proposed node classification method,this thesis selects enterprise safety production as a typical application scenario,through the design and implementation of enterprise safety risk assessment system,risk level classification assessment.Specifically,the evaluation module of the system constructs the data related to enterprise safety management as homogeneous graph data,trains the data based on the proposed algorithm model,and maps the training results to low-dimensional space for classification,so as to realize the functions of data extraction,processing and risk analysis.The test results show that the system can effectively assess the enterprise security risk,and the classification results can better accord with the enterprise reality,thus verifying the effectiveness of the proposed method.