Research On Enhancing Network Embedding Via Network Structure Learning

Network embedding,also known as Network Representation Learning,is widely used in social network,recommendation system,knowledge graph,natural language processing and other fields.Network embedding technology takes the information of network structure in application as input,and obtains the low-dimensional embeddings of network nodes.However,there are many defects in real data with network structure,such as lack of edges,containing error edges and so on,which make the effect of exist-ing models greatly reduced in application.At present,there are few researches on the network embedding for the data with defects in the network structure.In this thesis,we study the defects of network structure,such as edges missing,sparseness,errors and so on.An iterative network structure learning method is proposed to enhance the network structure and improve the effect of network embedding.Specifically,the main contents of this thesis include:1.For the problem of network structure defects such as edges missing,this thesis pro-poses an iterative network structure learning method,IAL.Each training round in the method is separated into several modules,includes subset selection,edge eval-uation,network structure update and network structure evaluation.This method can enhance the network structure data with edges missing,and improve the per-formance of network embedding.2.For the problem of network structure defects such as edges error,this thesis proposes a two step solution IAL-REC according to the characteristics of error edges.This method can effectively improve the quality of network structure with edges error and enhance the performance of network embedding.3.The methods are implemented as a practical tool.According to the characteristics of the methods,it separates and reorganizes some modules,including preprocess-ing,network structure learning,network embedding models,common algorithms and so on.Users can easily choose appropriate training strategies according to the characteristics of different tasks to improve development efficiency.4.Semi-supervised graph nodes classification experiments are carried out on common network structure datasets to verify the effectiveness of this method.The experi-mental results showed that these methods can effectively improve the quality of network structure data and the effect of network embedding model.