Dynamic Network Link Prediction With Multi-features

In the network,the link prediction task is to predict the state of the unobserved links based on the known network topology.In network link prediction,nodes and links change over time,and the formation of link relations between nodes is not only affected by the network structure of nodes,but also by the network evolution process,forming a complex dynamic network evolution process.In this context,link prediction is called dynamic network link prediction.Dynamic network link prediction is a common key problem in the application fields of network situation awareness,sequence recommendation and knowledge graph information completion,which has a very important research significance.Dynamic network data consists of links formed by nodes at different points in time.The traditional methods ignore the time series information of network evolution and regard the resulting topology as a static network research link prediction task.Recently,some studies set up time Windows to divide the dynamic network into different time slices,and predict the link state of the future time slice network according to the historical time slice information,taking into account the time sequence information of network structure change,but for the network within each time window,the timeliness of the link edge is not taken into account.In addition,the local structure information used by most dynamic network link prediction methods is not rich enough,and the link representation with rich features can be used to better judge the existing state of links.In order to improve the accuracy of link prediction,this paper designs a dynamic network link prediction model integrating multiple features.The main contributions of this paper are as follows:(1)A dynamic network link prediction model combining node representation sequence and local neighbor representation is proposed.In order to make better use of the time attribute of the network,this paper establishes the mechanism of node representation sequence.The node representation sequence mechanism introduced network representation technology and gated cyclic neural network to learn the low-dimensional representation vectors of nodes in different time slices.In order to make full use of the local neighbor characteristics of the link,a common neighbor representation of the target node pairs is established.Finally,the representation of the future time slice link not only includes the historical data features and the predicted new features,but also adds the local neighbor features of the node pair.Experimental results on five datasets show that the proposed model is superior to other benchmark methods in both AUC(Area Under Roc Curve)and AP(Average Precision).Furthermore,two network representation models and three classifiers are tested,and the results show that the proposed model is robust.(2)A dynamic network link prediction model based on link timeliness and local subgraph structure is proposed.When learning the temporal characteristics of the network,the temporal perceptual attention module is established to mine the temporal characteristics of the links in each time window.In order to enhance the representation of local structure features around links,this paper builds a local subgraph structure module based on the graph isomorphism algorithm(Palette Weisfeiler Lehman,PWL),and obtains the link expression containing the structure features of local subgraphs.Finally,the two features are combined to improve the performance of dynamic network link prediction.The experimental results show that the proposed fusion link timeliness and the structural features of local subgraphs significantly improve the accuracy of link prediction.The more time series features(the number of recent time slices),the higher the prediction accuracy;The number of neighbors in the local subgraph structure has little influence on the accuracy of link prediction when it changes within the range of 5-30.