Research On Link Prediction Based On Graph Neural Networks

A complex network is a discrete structure composed of nodes and intricate relationships between nodes.With the increasing development of science and technology,complex networks have become important and effective tools for people to model and analyze complex systems.However,when observing and measuring actual complex systems,the complex network data that people get are often incomplete or noisy.It is an important step in complex network analysis to determine whether there is an interaction or a link between two nodes,which has very important theoretical and practical significance.The goal of link prediction is determining which node pairs in a network will form new links in a period of time in the future or looking for links that should exist but are missing in the network.Link prediction is an essential and important step in processing and analysis of network structure data.Meanwhile,graph neural network is a powerful tool for processing network data.And it is gradually applied to the research of link prediction,which has attracted wide attentions from researchers.However,in these studies,the hierarchical feature of a network is ignored,and the importance of nodes is not evaluated.These are the problems that still exist in the current link prediction methods based on graph neural networks.Aiming at the above-mentioned issues of current link prediction methods,this thesis proposes link prediction methods that can effectively evaluate and distinguish the importance of nodes and make full use of the characteristics of the network hierarchy.The main contributions of this thesis are summarized as follows:(1)Aiming at the problem of making full use of the hierarchical structure of a network and distinguishing the importance of nodes,a subgraph hierarchy feature fusion model for link prediction is proposed.The model first maps the target node pair into a subgraph containing both the target node pair and its neighboring nodes.Based on the graph neural network model with graph pooling layers,nodes in the subgraph are aggregated hierarchically according to the importance of the nodes.The subgraph is thus represented by a single vector,which is finally used to judge the link existence of the target node pair by classification.(2)Aiming at problems such as the selection of neighboring nodes and the evaluation of node importance in link prediction,a link prediction model is proposed using weighted subgraph hierarchy feature fusion and attention mechanism.The model first transforms the original network into a weighted network through the re-weighting algorithm of edges.Then the model extracts a fixed number of subgraphs of a node on the weighted network,and adaptively learns the importance of nodes by introducing the graph pooling layer of attention mechanism.After that,the model aggregates the nodes features of the subgraph according to the learned nodes importance.Finally,a single vector representation of the subgraph is obtained,and fed into the classifier to judge whether there is a link between a target node pair.A large number of experiments have been carried out to verify and analyze the two link prediction models proposed in this thesis on a number of widely used network datasets.The experimental results show that the link prediction models based on subgraph classification proposed in this thesis have better link prediction performance on the tested datasets,compared with other current state-of-the-art link prediction models.The results demonstrate that the proposed link prediction models are effective.