Semi-supervised Node Classification Based On Graph Markov Convolution Neural Network

People live in an interconnected world.Many scenes in real life have the characteristics of graph structure.Every entity can be regarded as a node in the network.The attributes of an entity can be regarded as the characteristics of a node.The relationship between nodes can be regarded as the edge of graph structure,and graph is a good representation of information.In recent years,relational data modeling mainly follows two branches,being graph deep learning and statistical relational learning.Traditional neural networks do not consider the relationship between entities,they only use individual attributes to make prediction.Graph neural network can make use of both the attribute and structure of data,and statistical relation learning method can establish the dependency between tags,which can make the effect of learning prediction more powerful.The task of determining the label of a sample(represented as the label of a node)by looking at the labels of its neighbors is called node classification.Node classification has a wide range of application scenarios,including sematic segmentation in scene generated graph,human action recognition,recommendation system of social network,link prediction of macromolecules in biochemical field,etc.Following the work of Graph Markov Neural Network,this work aims to combine the advantages of graph deep learning for extracting more node feature information by using the structure and attributes of data and the advantages of statistical relationship learning for establishing node label dependency.This thesis explores a method to improve the accuracy of graph neural network node classification task.The method is applied to the condition where the input data model is a graph structured data model.In this work,Graph Markov Convolution Neural Network(GMCN)is proposed to train and test the node classification task data sets being Cora,PubMed and CiteSeer of graph data structure.Graph Markov Convolution Neural Network improves the operation of message propagation layer of graph Markov neural network(GMNN),and uses convolution operation to aggregate node information,which improves the accuracy of node classification task.The main improvement is that the GNN in GMNN for reasoning and learning uses recurrent GNN(RecGNN),while the GNN in GMCN for reasoning and learning uses convolutional GNN(ConvGNN).In this paper,we design experiments to verify the effectiveness of the proposed method and optimize the hyperparameters of the proposed graph Markov convolution neural network(GMCN)under the condition of semi supervised learning.We expect to introduce the efficient and fast propagation layer of graph convolution neural network(first order)into GMNN to achieve the goals of extracting representative node features and modeling label dependency,i.e.to construct a graph convolution neural network based on Markov random field.Therefore,the experimental design is mainly reflected in two aspects.On the one hand,by changing the propagation layer structure of graph neural network.On the other hand,the GMCN model is designed and constructed by optimizing the parameters.In the parameter optimization experiment,this work explores by changing the neural network optimizer,learn-ing rate and iteration times.In the comparative experiment,this thesis compares the results of several graph neural network methods and statistical relation learning methods which have achieved markable experimental performance in semi-supervised node classification task in recent years to prove the superiority of GMCN model.The neural network is trained under semi-supervised condition with three well-known datasets,being Cora,CiteSeer and PubMed.On Cora and CiteSeer datasets,GMCN model has higher classification accuracy than previous models.On PubMed dataset,GMCN model performs slightly worse than GMNN model.