The Extraction And Fusion Of Information In The Heterogeneous Network Representation Learning

Computer vision,data mining,social networks,community relations and many other fields of datasets and data relations in the real world can be intuitively represented by graphs/networks,and the need of these applications promote the related research and progress of graphs/networks,such as:node classification,link prediction,similarity analysis and so on.And the basis of all the research about networks is network representation,therefore network representation learning has attracted many attention in recent years.Most of the existing network representation learning methods focuse on the homogeneous networks.The homogeneous networks' nodes are in the same types,also the links in the homogeneous networks are in the same types.However,the relations on most of the real-world applications are complex,and homogeneous networks can not describe the differences between these entities and relationships,hence heterogeneous network representation learning is necessary to solve this problem.There are two problems in the existing heterogeneous network representation learning method.Firstly,most of the heterogeneous network representation learning methodd only utilize heterogeneity information and structure information,ignore the attribute information and the derived information from structure.The ignorance may lead to the loss of information.Secondly,the importances of these different information are different in the different applications,however there are few heterogeneous network representation learning methods on the importance of different information.The main contributions and innovations are following:·We proposed a biased random walk method.This method,on the one hand,captures the heterogeneity information by controlling the type sequences of the nodes.On the other hand,parameters controlled by nodes' rank capture the dervied information.We also proposed a heterogeneous network representation learning method based on attention mechanism and designed three masks to preserve the heterogeneity information,structure information and attribute information simultaneously.The results of node classification and clustering experiments has demonstrated the effectiveness of these two information extraction methods in heterogeneous representation learning.·We proposed two information fusion methods to fuse multiple network representations into one network representation.The first method is based on the autoencoder.By the designed objective function,different representations of the same node are gradually approaching in the low-dimensional space during encoding,hence different representations are fused.The seconde method is based on dynamic routing,which calculates the importance of different representations and fuses different representations.The results on node classification,clustering,visualization and similarity analysis show that our final representations obtain better test accuracy,and the ablation experiments show the effectiveness of our fusion methods.