Research On Coupling Perspective Based Network Representation Learning Algorithm

Network representation learning aims at mapping large-scale complex networks into a low-dimensional,real-valued and dense vector form,which can well capture various information of the network.Then,the representations of the network can be used into downstream tasks for network analysis.However,most traditional network representation learning algorithms rely on the connections between nodes of the network.The representation of each node is learned by aggregating the information of its neighbors iteratively,and the edge representation is also heavily dependent on the corresponding node.Actually,real networks are often sparse,and there are limited connections can be observed.The representations of nodes learned from this traditional information aggregation mechanism cannot effectively preserve the network structure,which further affects the representations of edges.In this thesis,the empirical analysis on real networks shows that the random walk in the line graph can produce sampling sequences with different node distribution from the original graph,which will bring more abundant structure information.Therefore,how to effectively integrate the different information from the original graph and the line graph is still a problem worth to be discussed.In addition,many nodes and edges in real networks contain rich attribute information.Attribute information can well supplement the deficiency of network structure information through effective fusion.However,due to the heterogeneity between structure information and attribute information,how to fully utilize node and edge attributes in the network is still a question to be solved.In view of the above problems,this thesis shifts its research focus from the original graph to the coupling perspective of original graph and line graph.Main work of this thesis is as follows:(1)To solve the problem of structural sparsity,this thesis proposes a network representation learning algorithm based on coupling perspective named CPBNE.This method first random walks in the line graph to obtain nodes sampling sequences which are different from the original graph,and then merges with the random walks in the original graph for coupled training.In this training mechanism,the representations of nodes and edges can be learned at the same time.The experimental results show that compared with the existing network representation learning algorithms,CPBNE can well preserve network topological information and achieve better results in downstream tasks.In sparse networks,CPBNE has good scalability,and its execution efficiency is more than three times higher than the existing deep learning-based methods.(2)In order to fully use the information in the network and solve the problem of fusing node attributes and edge attributes with structural information,this thesis further improves the lack of nodes and edges attribute information in the random walk sequence.An attribute-enhanced network representation learning algorithm based on coupling perspective(AE-CPBNE)is proposed.The attributes of nodes and edges in the network are constructed as new attribute nodes.Through the transformation mechanism of line graph,edge attributes in the original graph can be naturally converted into the node attributes in the line graph.Then,the attribute contexts with distance attenuation are generated which will be fused with the structure contexts in the process of coupled training.This thesis verifies the performance of AE-CPBNE in two real attribute networks.Compared with traditional attributed network representation learning algorithm,the performance of AE-CPBNE keeps in top two in node classification,edge classification,link prediction and link reconstruction tasks.