Research On Link Prediction Method Based On Deep Learning Of Evolutionary Features

Networks are very effective tools in the modeling of complex systems composed of interacting elements.In scientific research,network data mining has a wide range of applications in many fields,including protein interaction networks,social networks,transportation networks,and telecommunications networks.The problem of predicting new relationships that may appear in the network is also called link prediction.Link prediction aims to infer the behavior of the network link formation process by predicting missing or future relationships based on currently observed connections.In this chapter,we will obtain the current status and advantages and disadvantages of link prediction algorithms by reviewing and analyzing the general technology for dealing with link prediction problems.Generally speaking,link prediction problems can be divided into static prediction problems and dynamic prediction problems.The former considers using a snapshot of a network with a fixed structure to predict links,while the latter solves this problem based on a snapshot stream of a dynamic network.But most of the current networks are dynamic networks.The number of nodes and topology of the network are constantly evolving over time,and static link prediction cannot capture the characteristics of the network evolving over time,so the accuracy of the prediction results is not particularly ideal.In order to improve the accuracy of link prediction,this paper focuses on the characteristics of the network evolution over time,and on the basis of studying the time window division,through the ternary closed network representation learning model,quantified the development of open triples into closed ternary The probability of the group captures the evolutionary structural properties of the network to learn the embedding vector of the vertices,and finally uses the distance of the embedding vector of the node to predict the result of similarity measurement.The main contents of this thesis are as follows:(1)Selection of time window size.At present,most networks evolve over time.If you just divide the time slices arbitrarily,you may not pay attention to the evolution information of the network,and it will affect the accuracy of link prediction.Then,in view of the frequent changes of the dynamic network topology,the network is divided into a suitable time window in chronological order.In this experiment,the SOTS algorithm is used to divide the time window,which can quickly reduce the loss of network information.Accurately divide the time window of a suitable size,and the data set under this time window can better display the information of the dynamic network.(2)The ternary closed model performs network embedding representation learning.The appropriate time window size is selected through the time slice division method,and the dynamic network data set under the corresponding time window is captured by the three-element closed dynamic network embedding representation learning algorithm to capture the evolution structure attributes of the network to learn the embedding vector of the vertex,the three element closed The model quantifies the probability of open triples developing into closed triples,so as to learn the embedding vector matrix of each vertex at different time points.As the smallest local structure that maintains the network topology,the triplet has the characteristics of structural balance and stability.The change of the triplet closure is the cause of the change of the link connection.This paper uses the network representation learning method of the ternary closed model to establish a link prediction model.(3)Build a link prediction model.The obtained embedding vector matrix is used to obtain the Euclidean distance between the node vectors through the similarity index,the node similarity matrix is obtained,and the possibility of link prediction is calculated.Experiment with the constructed link prediction model on the data set,verify the prediction results through predictive indicators such as AUC,and compare it with the prediction results of node2 vec and TNE algorithms.It is concluded that the link prediction model proposed in the article has good prediction accuracy.Significantly improved.