Research On Link Prediction Task Based On Attention Network

After entering the 21 st century,information technology represented by the Internet has developed rapidly,and more and more complex networks have appeared in daily life,such as online social networks based on social platforms such as We Chat and QQ.Therefore,it is of great significance to study and explore the dynamic evolution mechanism behind various networks.With the deepening of research,more and more network characteristics have been discovered,and the link status in the network is an important indicator of network properties,and the link prediction problem has attracted attention.The role of link prediction is to supplement possible links when the network structure information is incomplete.For static networks,it can help people find missing links;for dynamic networks,it can help people predict network topology changes in the future and understand the evolution mechanism of the network.Therefore,link prediction has potential great value in solving practical industrial problems,such as friend recommendations applied to social software or product recommendation in online shopping malls.However,in the process of network evolution,the interaction between nodes is not only binary,there are group interaction behaviors on many occasions.Therefore,it is of great significance to study the high-order potential interaction relationship between nodes to explore the evolution mechanism of the network.This paper mainly focuses on building a graph neural network(attention network)model of high-level structural information,and then explores the network evolution mechanism and predicts and regulates the network structure.We choose the attention network as the main research method because it can effectively extract node feature vectors while taking into account the correlation and difference between each node and a node and its neighbors,which is its technical advantage.At the same time,a multi-head attention mechanism is set in the graph attention network,and the feature components of different network evolution directions are automatically learned by using this mechanism.By aggregating these components,the task robustness of the model can be effectively guaranteed to achieve the best results.This paper is mainly divided into the following parts:1.First of all,combined with the formation and development of the research problem,the background,purpose,and significance of the topic selection of this paper are clarified.Second,the existing link prediction models are summarized.This paper classifies and summarizes the link prediction models of complex networks,and analyzes the advantages and disadvantages of various models according to the dynamic and static states of the network.Then,the research content of this paper and the organizational structure of the article are briefly introduced.2.According to the dynamic and static state of the network,the link prediction problem in each state is described.The relevant technical methods used in the research work are introduced in detail,including graph neural networks based on frequency domain and spatial domain,attention mechanism,graph attention network,etc.3.A link prediction neural network model for static networks is proposed.The learning of network structure characteristics depends on the balance between local and global structures,and existing models are often achieved by adjusting hyperparameters.However,if the model can learn effective local and global information more intelligently,the performance can be improved and further improved.In order to realize the conjecture,we propose GANat,a static link prediction model that relies on Graph Attention Network(GAT).The model first uses GAT to capture the local structure information of the first-order neighbors of the node.Then use the Attention Network(Attention Network)to extract the effective characteristics of multi-order neighbors to expand the structural characteristics of network nodes.Finally,the combined vector of node pairs is fed into the logical regression model to obtain the probability prediction value of the link between node pairs.Experiments in Facebook,Twitter,and other data sets show that the overall performance of GANat is better than that of the baseline model,and it has certain advantages in balancing local and global structural information.The model also effectively solves the time parallelism defect of repeated convolutional neural networks.4.A graph neural network model(GRL＿En SAT)is proposed for dynamic link prediction tasks.Generally,the learning of time patterns follows a timeliness hypothesis: old information is not as important as recent information.However,the timeliness hypothesis is not always true.Therefore,consider adding more flexibility when weighing historical data.The model first learns the local and global structure information of the nodes at each time step of the network through GANat.Secondly,the Temporal Attention Network(Temporary Attention Network),which is improved from the attention network and has the ability to independently determine the importance of time series information,is selected to capture the dynamic evolution characteristics of the network.Since the evolution of the network is affected by a variety of potential factors,in order to improve the robustness of the model results,we apply an independent multi-head mechanism to all attention networks,so that the model has the ability to learn evolutionary characteristics from multiple evolutionary angles.Then,the regression model is used to obtain the final prediction probability value.The experimental results in UCI,Department,and other data sets show that GRL＿En SAT has significant advantages in performance robustness,and its indicators are better than those of Dy AERNN,Dy SAT,and other models.5.The main work of this paper is summarized,and future research work is prospected according to the harvest in the research process.At the same time,the future research direction of the link prediction problem is partially conjectured and described.