Network Embedding Based On Information Enhancement

Network is a kind of data structure to describe the relationship among nodes.It can be used to define complex data relationships,which can be scaled to extremely huge database.Therefore,it is suitable for creating data models under the setting of big data,and gains people's favor.Common applications on networks include social network-s,citation networks,word networks,knowledge graphs,chemical substance networks and protein networks.The goal of network representation learning is to extract com-plex and representative features from the networks,so as to support the downstream knowledge discovery and knowledge mining tasks.In the recent years,with the pop-ularization of big data and the progress of deep learning technology,the theory and methods of network representation learning have made a breakthrough development.However,there are still some challenges on the utilization of network informa-tion.Firstly,in order to better explore the information of networks,researchers often use random walk based sampling strategy to expand the neighborhoods,in order to p-reserve network structure information.However,the strategy of random walk is based on experience.The amount of information are not defined,which brings too much randomness.Secondly,attribute network provides rich attribute information.How-ever,the existing attribute network learning methods can only utilize the attributes of the target node,but ignores the context attributes.Thirdly,in attributed networks,it is a practical requirement to represent different attributes and composite attributes.The existing methods often represent each attribute separately,which is difficult to utilize the relationship between different attributes,and the utilization of attribute relationship information is not enough.Last,in the settings of semi-supervised learning,graph neural network efficiently aggregates the local information for producing good node representations.However,due to the over-smoothing problem,it is difficult to make use of the global information of the network.This thesis aims to solve the above limitations in network representation learning,and enhance the information utilization ability of each step.The main work includes:1.The current random walk strategies are mainly based on experience.To solve the problem,the weighted information entropy is designed to measure the informa?tion amount of the random walk sampling results.It is found that the higher the information amount,the better the performance of the subsequent classification and link prediction tasks.Based on the discovery,a sampling method based on maximizing weighted information entropy is proposed.Experiments show that the sampling method proposed can effectively improve the accuracy of subse-quent tasks.2.The existing attributed network embedding methods in unsupervised learning cannot model the context attributes.To solve the problem,two objective func-tions are proposed,one for preserving the structure features of the context nodes,and the other for capturing the context attributes.Therefore,our methods can yield node representations with more information.Experiments show that the proposed method is significantly better than the benchmark methods,especially when the data is sparse.3.It is difficult to utilize the relationship among different attributes.The thesis solves the problem within the settings of citation networks.The thesis proposes a encoding-decoding model based on recurrent neural network,which first maps different scholar attributes,such as title,author and journal,to the same vector space,and then tries to maximize the likelihood of the citation relationships a-mong attributes.Hence,our method can model the citation relationships among different attributes and support some interesting applications,such as reviewer recommendation,journal publication recommendation and so on.Moreover,due to the utilization of the relationships among different attributes,this paper also significantly outperforms the benchmark methods on the task of paper similarity measurement.4.To solve the problem that Graph Neural Networks cannot model global informa-tion.In the framework semi-supervised network embedding,a graph neural net-work is proposed to learn a model which can utilize global network features.The model first uses unsupervised random walk methods to learn the global structure and attribute features of nodes from the network.Then,a global graph neural network algorithm is proposed to generate the final feature representation.The thesis studies the problems of the plain network representation and attribute net-work representation.The experimental results show that the proposed method can significantly outperforms the benchmark methods in both settings.