Study Of Graph Embedding Learning Methods Based On Deep Neural Networks

With the widespread development of information technology,it is becoming increasingly popular to use complex networks to model the relationships between entities of complex systems.For example,social networks in social media,citation networks in academia.Graph analysis has aroused considerable research interests.Graph analysis tasks can be classified into two categories: node-level tasks such as node classification,link prediction and node clustering etc.,graph-level tasks such as graph classification.Most machine learning-based network analysis algorithms heavily rely on the node representations and graph representations.However,traditional graph representation methods such as adjacency matrix fail to reflect more complex underlying structure relationships,and suffer from the problems of high time complexity with the graph scale increasing.Therefore,graph embedding aiming to encode graph into low-dimensional space while preserving network structure information and inherited properties has attracted increasing attention.This paper studied graph embedding learning including node embedding learning and whole graph embedding learning,and proposed deep neural network-based graph embedding methods to learn node embedding or whole graph embedding address the node-level or graph-level tasks.In summary,the main contributions of this thesis are as follows.1)This paper proposed a semi-supervised node embedding model named SNE to integrate structural information,text features and category attributes into node embedding vectors simultaneously.Most existing node embedding methods only focus on preserving structure information and ignore node features and inherited attributes.Moreover,those methods fail to model the highly non-linear structure information of networks,due to the use of shallow models.To address the issues,the proposed method designs a structure preserving module to measure structural similarities between nodes and a supervised loss to ensure nodes in the same class have more similar embedding vectors,and utilizes the stacked auto-encoder network to model the highly non-linear information.Extensive experiments on real-world datasets demonstrate that the proposed method is superior to the state-of-the-art baselines in a variety of tasks,including node visualization,nod classification and node clustering.2)This paper proposed a multi-level coarsening-based graph convolutional network(MLCGCN)method to learn graph representations and address graph classification tasks.To solve the problems of over-smoothing,lack of hierarchy and interpretability of graph convolutional networks,we first develop new insights into the convolutional architecture of image classification from the perspective of graph analysis,and then present the two-stage MLCGCN architecture for graph classifications.In the architecture,we first introduce an adaptive structural coarsening module to produce a series of coarsened graphs of different scales and then construct the convolutional network based on these graphs to learn the whole graph embeddings.The proposed method has the advantages of learning graph representations at multiple levels while preserving the local and global information of graphs.Experimental results on multiple benchmark datasets demonstrate that the proposed MLC-GCN method is competitive with the state-of-the-art graph classification methods.