Research On Graph Neural Network Based Network Attribute Prediction Algorithms

In the real world,there are a lot of network structure data,the network structure have attributes,which include node information,edge information and network global information,etc.Among them,the information of nodes will be used to describe the characteristics and status of nodes,such as the information of the user in the social network including the user’s identity information,hobbies,etc.The edge information will be used to describe the relationship between two nodes,such as the intimacy between two users or the attractiveness between users and products.These information are the attributes of nodes and the attributes of edges,which are collectively called network attributes.By studying and mastering network attributes,we can perform different tasks to obtain potential information about the network.However,network attributes are often incomplete,which will lead to difficulty in our understanding of the network,which is not conducive to mining the potential information of the network.Aiming at the missing network attributes,manually recovering the attributes will cause a lot of cost loss,especially not suitable for large-scale networks.Network attribute prediction is to predict unknown attributes in the network using prediction algorithms based on the known attributes in the network.The purpose is to reduce the loss of various costs,improve the accuracy of unknown attribute prediction and obtain the complete attributes of the network.This thesis focuses focus on two sub-problems of the network attribute prediction problem,namely the prediction of network link relationships and the prediction of missing attributes of network nodes.These belong to the prediction of edge information and node information in the network attributes respectively.Link prediction is to predict the relationship between unknown node pairs through known network attributes and node missing attribute prediction is based on known network attributes to predict the unknown part of the node attributes.To address these issues,this thesis proposes a network link relationship prediction algorithm based on graph neural network and a network node missing attribute prediction algorithm based on graph neural network to solve these two sub-problems.Firstly,we propose a Dual-Graphlrp model framework based on dual graph neural network to obtain the spatial structure information and attribute information of the graph nodes through the message passing mechanism and learn the embedding vector about the node and construct a relationship extractor for the combination of the embedding vector of the two-two graph nodes to learn the relationship between the two nodes.Secondly,we propose a node missing attribute prediction model based on graph autoencoder that uses the learned graph structure.Through our proposed graph structure learning algorithm and graph sampling algorithm,a sparse graph structure that can reflect the correlation of nodes is obtained.The process of encoding and decoding is reconstructed to obtain the attributes of the nodes.Experimental verification in different scenarios shows that our proposed algorithm can effectively predict the link relationship of the network and the missing attributes of the nodes.It proves that the combination of network structure information,network local correlation information and deep learning model can effectively predict network attributes.