Research On Attributed Network Representation Learning With Multi-Level Structure

Complex networks are becoming everywhere in a large number of practical applications in the form of social network,citation network,biological network or others.Analyzing complex networks plays a vital role in many applications and disciplines,such as node classification,link prediction and so on.However,large-scale network analysis is still a challenge.In recent years,Network Representation Learning(NRL),as a new learning method,can embed network nodes into low-dimensional vector space,and has been widely applied to network analysis tasks.Attributed network is a network in which nodes or edges contain attribute information.Most of the existing methods are based on the network structure,and do not pay attention to the attribute information that can provide richer content for network representation.However,due to the heterogeneity of structure and attribute information,it is difficult to establish a joint learning representation.In addition,the network structure has multiple levels of complexity,which makes it difficult for existing methods to capture the complete network structure.In response to these problems,this thesis explores a network representation learning method that can maintain multi-level structure and attribute information at the same time.The specific work contents are as follows:(1)In order to effectively integrate network structure and attribute information,this thesis proposes a community aware and relational attention method for attributed network representation learning(CARA),which retains rich structural information while using attribute information to model the relationship between nodes to enhance network representation.Specifically,the community aware random walks use network characteristics to expand the neighborhood structure to enhance the structural representation,and make up for the incomplete network embedding caused by network sparsity;the relationship attention mechanism models the relationship between nodes,and it is used for different neighbors to learn discriminative attribute representations.(2)Aiming at the problem that the existing methods retain the local structure information of the network and cannot capture the long-distance high-order proximity and network structure characteristics,this thesis further proposes a graph compression attributed network representation learning method GCANRL,to enhance the existing network representation learning method.First,a compression strategy is proposed,which can compress the original attributed network into a series of hierarchical graphs from fine to coarse according to the characteristics of the network structure.The hierarchical graphs can reveal the multi-level structural characteristics and attribute information of the network.Then using these hierarchical graphs to perform hierarchical network representation learning,so that the final node representation contains both fine-grained and coarse-grained information.This thesis first proposes a method for joint representation of the network structure and attributes,which effectively realizes the attributed network representation learning;then the proposed graph compression strategy is used to maintain the network multi-level structure and attribute information,and the experimental results verify that GCANRL has a further effect on CARA and other NRL methods,so it has good theoretical and application value.