Research On Link Prediction Algorithm Of Attention Flow Network Based On GNN

As an important branch of complex network and data mining,link prediction aims to mine the hidden potential law of network by analyzing the known information of network nodes and network structure,etc.,and predict the possibility of any two nodes in the network to generate edge connection in the future.Link prediction reveals the basic mechanism and evolution law of the network from different angles,which has important practical significance in alleviating information overload,providing personalized service and assisting users to make decisions.However,link prediction methods still have the following problems: On the one hand,traditional heuristic methods only learn the graph structure features of the network,and lack comprehensive application of explicit features and implicit feature information,and most methods are based on undirected powerless networks.On the other hand,the current link prediction method is mainly based on low-order local structure characteristics,ignores the high-order community relationship between nodes,and does not consider the combination of node embedding and community embedding to improve the performance of link prediction.To solve the above problems,this thesis constructs an attention flow network based on online user clickstream data,and proposes an attention flow network link prediction algorithm based on GNN.The research content of this thesis includes the following three aspects.Firstly,we proposes a feature learning algorithm AFN＿FL based on the R-GCN algorithm and attention mechanism.First of all,we abstracted different edge directions between nodes in the attention flow network into two different edge relationship types and used the attention mechanism to aggregate node attributes and edge attributes.Secondly,the R-GCN algorithm learned the graph structure features,explicit features,and implicit features of the network.Finally,the probability of whether the edge relation between nodes is valid or not valid is obtained by the feature vector through the scoring function,so as to finish the AFN＿FL algorithm.Compared with GCN,GAT,and other algorithms,AFN＿FL improves ACC value by 3.52%,accuracy by 3.2%,Recall value by1.59%,and F1 value by 3.28%.Secondly,we proposes a GNN-based attention flow network link prediction algorithm NCELP,which combines low-order local structure and high-order community relationships.Firstly,we used the DRNL algorithm and Louvain community detection algorithm to explore the structural importance and community structure of the attention flow network,and endowed each node with unique node labels and community labels as explicit features.Secondly,we used the AFN＿FL algorithm to learn local features,multivariate Gaussian distribution to learn community topology information,and integrated node embedding and community embedding by the DGCNN method as implicit features.Finally,the link prediction problem is transformed into a binary classification problem based on the graph structure characteristics to realize link prediction.Compared with GIC and NPGNN,the AUC and AP value of NCELP are increased by 9.83% and 1.49%.Thirdly,application research of link prediction algorithm in microblog rumor propagation data.Firstly,we selected ten rumor microblogs based on the public rumor data and extracted the information on rumor microblogs and related microblog users.Containing 47,086 users and 525,411 forwarding/comment relationships,were used to construct the microblog rumor propagation network.Secondly,we transformed the prediction problem of the microblog rumor propagation path into a link prediction problem,and used the link prediction algorithm in this thesis to predict the rumor propagation path.Finally,we verified the applicability and effectiveness of the link prediction algorithm in the rumor data through experiments.