Research On Link Prediction Based On Network Representation Learning

As a mathematical model to characterize the connections and interaction patterns between various parts of the system,complex networks have been widely used in many disciplines.As a powerful and general data mining analysis tool,complex networks are always expected to be complete and accurate,but the information collected in the real world is always not comprehensive and contains noise,and the link prediction problem emerges as the times require.Link prediction,which attempts to predict whether there are missing or possible future links in a network,is an important and challenging problem in complex network analysis.In recent years,link prediction methods based on network representation learning on subgraphs have shown promising performance.However,there are still some issues to be resolved.First,the capture of feature information contained in a single local topology attribute in a subgraph is not comprehensive.Second,insufficient extraction of global structural information of the network and low model training efficiency.Aiming at the above problems,this thesis proposes a model that can fully capture the structural information of subgraphs and a model that effectively utilizes more global structural information and achieves efficient training.The main contributions of this thesis are as follows:(1)For the problem that the structural feature information of the subgraph is not fully captured,a link prediction method for multi-level supervision on line graph is proposed.The model first extracts its associated subgraph around the target node pair in the network and then transforms this subgraph into a line graph,where the target node of the line graph corresponds to the target node pair to be predicted in the original network.This method analyzes multiple local topological properties of the associated subgraph for constructing line graph features as input features of a graph neural network,and an intermediate supervision mechanism is used to guide feature extraction of the graph neural network.Finally,the classification feature of the target node is input into the classifier for binary classification,which is to predict whether the corresponding node pair in the original network has a link.(2)For the problem that the global structure information of the network is not sufficiently extracted and the model training efficiency is low,a link prediction method by node classification on line graph is proposed.The method also uses line graph transformation to cast the link prediction problem as a node classification problem.This method first randomly selects links that do not exist in the network as negative instances,transforms the input network containing positive and negative instances(links and non-links in the original network)into its line graph,and then extracts the subgraph around a given target node on the line graph,where the target nodes of the subgraph correspond to the target link to be predicted in the input network.It constructs the subgraph features with global structural information and fuses these features through simple graph convolution operations to learn the classification feature of the target node.Finally,the extracted feature of the target node is fed into a binary classifier to predict the existence of the target link in the input network.Multiple types of dataset networks are selected for a large number of experiments to verify and analyze the two proposed methods.The experimental analysis shows that the two methods proposed in this thesis can outperform most of the baseline methods in link prediction performance,and also verify that the second method can improve training efficiency and guarantee the performance.