Research On Network Representation Learning Method Based On Edges And Attributes

In the past decade,the development of information science and technology and mobile Internet has given birth to a huge amount of data,which and their relationships are often abstracted as networks,such as social relationship networks,paper citation networks and bioinformation networks.Naturally,learning effective information from these information networks has become an urgent practical demand.Therefore,network representation learning technology has received widespread attention.So far,researchers have proposed many methods for network representation learning.generally speaking,there are two most representative methods,one is based on random walk and shallow neural network represented by deepwalk algorithm,and the other is based on deep neural network represented by SDNE model.This paper will discuss and study based on the existing research results and the existing problems,the main research contents are as follows:First,in order to solve the problem that most of the existing algorithms are devoted to the study of network topology and ignore the edge information in the network,an edge constraint based network representation(ECNR)algorithm is proposed.The algorithm constrains the random walk by the information of network edges.That is to say,we first learn the representation of connected edges via node representation learning method,then use constrained random walk in the process of node sampling,forcing each sampling to choose the node with stronger edge association and more similar edge vector as far as possible,and finally use the word vector learning model skip-gram to train the node sequence and get the vector representation of the node.Experiments are carried out on three datasets Cora,Citeseer and wiki.Compared with the comparison algorithm,the ECNR algorithm achieves better results.Secondly,aiming at the problem that the existing algorithms ignore the differences of attributes between nodes and their neighbors,a weighted early fusion representation learning based on autoencoder(AE-WEF)is proposed.In the early stage of the algorithm,the attribute features and structure features of the nodes are weighted and fused,and then the fused features are input to the autoencoder for optimization,and the final node vector representation is obtained.The AE-WEF model is applied to the node classification task of data sets Cora,Citeseer and Wiki.Compared with the comparison algorithm,the average improvement of Macro-F1 and micro-F1 is 0.1 and 0.103 respectively.