Research On Link Prediction In Dynamic Network Based On Deep Learning

The complex relationship between the real world can be expressed by the complex network model.For example,various interpersonal relationships,road traffic,complex interaction between cells and the structure of links between the Internet,etc.With the development of complex network research,many important network characteristics are gradually discovered,and as an important indicator of network characteristics,the link connection in the network is the most important indicator.The link prediction problem in complex networks is applied.The link prediction problem in network is to predict the possibility of connection between two nodes in the network that have not yet produced a connection.This prediction includes both the prediction of unknown connections(links in the network that are actually present but not detected by us),but also the prediction of future links(links that are not present in the network but should exist or are likely to exist in the future).Link prediction has great value in the theoretical research of complex networks.In the static network environment,it can help people to find the possible missing links and improve the network structure information.In the dynamic network environment,link prediction can help people predict the network topology structure and the mechanism of network evolution for the next period.Each network evolution mechanism may contain an accurate link prediction method,and each excellent link prediction method may also reveal a network evolution mechanism.This paper starts with the research of static network and the link prediction of dynamic network.It puts forward a method of static link prediction using SMTV cbow method in the basic static network environment.In the dynamic network environment,a link prediction method based on the convolution long-term memory network is proposed.In order to effectively and accurately mine the relationship between the self-attributes of node and structure of network and apply it to link prediction,this paper is inspired by probabilistic language retrieval research,we propose a link prediction method based on CBOW model.By using the node sequence library which containing node neighbor information and network connectivity information to train the CBOW model to generate the node vector.We proposed a new similarity evaluation index,based on the self-attribute of the node vector and the tendency between the node pairs,called self-measurement tendency of vector(SMTV).Using this similarity index for network link prediction.We experimented on three real data sets of PPI-Yeast,Facebook and Power Grid,and obtained the AUC values of CN,AA,LP and Node2 vec respectively.Compared with the method of lowest AUC in above four methods,the CBOW-SMTV method has an increase of 5.3109%,14.4955%,and 41.9747% respectively.And even compared the method of highest AUC method,our method has an increase of 0.2497%,0.6921%,and 9.54114%.Therefore,the link prediction method based on CBOW-SMTV can effectively combine node attributes and network structure information to improve link prediction effectiveness.In the dynamic network environment,considering the time sequence characteristics of network evolution,combining with the representation of image information in computer,the network time series is divided into several continuous time windows according to the average connection time.By using the network connection in each time window,the connection frequency matrix including the number of connections in time slice and the network connection are constructed.The long connection time matrix and the active matrix considering the number of active nodes are considered.The state diagram in each time slice is used as input to train Conv LSTM network,and the adjacent information and timing information of nodes are extracted to obtain more accurate prediction results.The AUC values obtained by experiments in infocomm05,MIT and Email-Eu-core are better than those of CN,AA and Katz.It is proved that AG-Convlstm algorithm can predict the network connection in the next time slice more accurately.