Research On Representation Learning And Link Prediction Under Attention Flow Network

The attention flow network is an important branch of the complex network field,meanwhile network representation learning is the essential way to reasonably express the node attributes and structural information within network.Therefore the representation learning of attention flow network will effectively express the network information,but also has wide range of application scenarios.As one of those scenarios,link prediction mainly use known nodes,network structure and other information to predict potential connections between nodes and proceed to resolve the missing connection problems within network.Not only recognizing the online users’ preference,but also the link prediction research in the attention flow network has the ability to make predictions on the users’ online click behavior.Although very few studies on it were published,this paper explores the problem of representation learning and link prediction based on attention flow networks and the specific research content are as follow:Firstly,The effective representation of the attention flow network is achieved from two perspectives,network structure and matrix decomposition.First,from network structure,the representative learning on network structure can learn the network information of the local neighborhood,but also obtain the network information and proceed to reasonably represent the network by using skip-gram of natural language processing.And as for matrix decomposition,the similarity matrix of the network is factorized to obtain the original embedding of the initial node,and the global and local information are merged by spectral propagation to get a reasonable representation vector.A large number of experimental results show that characterizing the attention flow network based on structural features can effectively obtain the structural features of the network,which is more comprehensive than other representation learning algorithms.The representation learning of the network based on matrix decomposition can not only effectively filter out "noise",and essentially optimize the network structure to obtain better representation vectors.Secondly,This thesis proposed the SMLP algorithm,a link prediction algorithm based on structural features.First,obtain the node neighbors in random neighborhood of the node in the attention flow network,and perform a reasonable order;second,for any two nodes,use the node neighbor sequence under the same neighborhood to calculate the similarity of the two nodes;then,construct a multi-layer complete graph.Any nodes between each layer of the complete graph are connected to each other,and the weight of the connection is the structural similarity between the nodes,and the complete graph of each layer is connected correspondingly;finally,the strategy of weighing the weight is adopted to migrate and obtain the representation vector of the network,and use the better fault tolerance and memory ability of the deep neural network to predict the online click behavior of collective users.We have massive experimental results to prove that,for the link prediction problem of attention flow network,the AUC value of the SMLP algorithm is about 95.2%,and the algorithm is generally good in accuracy,precision and F1 value.Thirdly,A link prediction algorithm Ada-PE based on matrix factorization is proposed.The algorithm builds a similarity matrix based on the attention flow network,transforms the network representation problem into a sparse matrix decomposition problem,and uses randomized tSVD to achieve the initial graph embedding of the network.Considering that general graph embedding can only learn local structural information,this paper integrates the global information of the network further through spectral propagation,combines the global clustering information and local smoothing information of the network into the representation vector of the network,and then applied it to the link prediction of characterization learning,so as to achieve excellent results.Our experimental data shows that,for the link prediction problem of the attention flow network,the AUC value of the Ada-PE algorithm is about 97.6%,and the other evaluation indicators have increased by about 8.13%.