Research On Network Representation Learning Method

In recent years,various online social network applications have appeared in people's life and have produced a large amount of network data sets.Traditional network representa-tion methods cannot deal with these large-scale network data efficiently and quickly.With the development of deep learning technology and inspired by the idea of word embedding in the field of natural language processing,automated network representation learning has become a new research hotspot.Network representation learning aims to project nodes in the network into a low-dimensional dense vector space through deep learning methods.Fur-thermore,it can be applied to tasks such as network visualization,node classification,link prediction,and community discovery.Network data can be divided into undirected networks,directed networks,and knowledge graphs according to the complexity of the structure.The main work of this paper is to propose methods to improve the performance of network rep-resentation learning in different types of networks.The main contents include:(1)Undirected network representation learning method based on gated graph attention mechanism.Considering that the feature weights of the attention coefficients between nodes in graph attention network are fixed and the model cannot sense the network structure,we propose a method for improving the graph attention network using gated mechanism.This method inputs the neighborhood information of the nodes into the gating unit,so that the feature weights can be flexibly adjusted when calculating the attention coefficient.Experi-mental results show that the model achieves good results on the undirected network repre-sentation learning task.Compared with the benchmark model,the mean accuracy of trans-ductive task has been improved by 0.5%on Cora and the micro-F1 value of inductive task has been improved by 1.3%on PPI.(2)Directed network representation learning method based on hierarchical information.Because the ring in the directed network destroys the asymmetric transitivity and makes the model difficult to learn the global structural features,we propose a method to introduce the hierarchical information of the nodes into the network representation learning.Experimental results show that the model has a significant improvement over the benchmark model in the directed network representation learning task.Compared with the benchmark model,the AUC value of link prediction has been improved by 7.5%,1.1%,4.3%and 4.6%on Wiki-Vote,Jdk-dependency,Cora-citation and Cit-HepPh.(3)Knowledge graph representation learning method based on improved vector projec-tion distance.One-to-many,many-to-one,and many-to-many complex relationships in the knowledge graph are difficult to learn.We introduce an adaptive matrix and adjust the weights in the loss function on the knowledge graph representation method based on the vector projection distance.Experimental results show that the model has a significant im-provement over the benchmark model in the knowledge graph representation learning task.Compared with the benchmark model,the hits@10 value of link prediction has been im-proved by 8.6%on FB15k.This thesis analyzes the structural fearures of undirected networks,directed networks,and knowledge graphs,and explores the problems in representation learning of these net-works and finally proposes corresponding solutions.The research on network representation learning helps lay the foundation for further research and use of large-scale network data.