| Information network is an abstract form of the real world which can associate various objects and express the interaction between objects.Therefore,network analysis has become a research hotspot in both industrial and academic circles.Network representation learning aims to realize an algorithm to transform high-dimensional network nodes into low-dimensional dense continuous vectors which can be used in subsequent application tasks.At present,there are many researches on network representation learning methods at home and abroad.The existing research mainly focuses on dimensionality reduction and matrix decomposition methods,such as LLE,LE,GF and so on.In addition,the existing research is mainly aimed at homogeneous networks with the same node type and relationship type,which is not applicable to heterogeneous information networks with multiple types of objects and link relationships.In the real society,a large number of abstract heterogeneous information networks exist,such as thesis network and circle of friends network,including various complex network topology information and rich node attribute information.However,the existing research on network representation learning methods is often difficult to fully consider the node attribute information,or only consider the attribute information and ignore the node topology.It can not completely maintain the semantics of the whole network.To settle the mentioned problems,this paper studies a network representation learning method based on word embedding,taking into account the structural characteristics of heterogeneous networks and the attribute characteristics of nodes in the network.The specific work is shown as follows :Firstly,this paper proposes a network representation learning method based on word embedding.The word representation learning model in the area of natural language processing can maintain the semantics of words and realize the representation of words.In this study,the idea of word embedding is introduced into network representation learning.The definition of related concepts and the formal description of the problem are given.This paper gives a minute description of the basic thought and specific procedures of the method,and analyzes the spatiotemporal complexity of the method.In addition,this study extends the application of network representation learning in homogeneous networks to heterogeneous networks,and proposes an improved node sampling method to capture the original attribute heterogeneous network topology information more truly.By decomposing the original heterogeneous network structure,AHIN is decomposed into isomorphic network by meta path processing,and an attribute bipartite graph is decomposed by attribute set.On the bipartite graph,the final representation vector is obtained by improved node sampling method;It is proposed to set the termination probability parameter,so as to obtain node sequences of different lengths in the process of node sampling,so as to avoid cyclic sampling of visited nodes;In addition,the node influence parameters are set,and the nodes with high influence value are sampled first to fully restore the original information of the network;Finally,the improved node sampling sequence is trained by skip gram embedding model to obtain the vector representation of nodes.Finally,we carry out the task of multi label classification on three real network data sets:Wiki,IMDB and DBLP.The feasibility and efficiency of the proposed method are verified by discussing and analyzing the results of comparative experiments. |
