Research On Adaptive Network Embedding

Network embedding,also called network representation learning,aims at learning lowdimensional vectorial representations for nodes in networks.The learned node representations can be used for various network analysis tasks,such as node classification,link prediction,node clustering,network alignment,and so on.In recent years,with large amount of network data being produced,network embedding has been attracting increasing interest.According to the number of used networks during the process of node representations learning,network embedding can be generally divided into two categories: network embedding in single-network scenario and network embedding in multi-network scenario.Although many researchers have studied network embedding in single-network scenario and network embedding in multi-network scenario,there are still some problems have not been solved.For network embedding in single-network scenario,the existing network embedding works usually capture only one of two properties of nodes,i.e.homophily and structural equivalence,however,in real world networks,nodes usually exhibit these two properties at the same time.For network embedding in multi-network scenario,the existing works usually ignore that the influence of nodes will change with the direction of edges in directed networks.In addition,the existing works often utilize domain adaptation techniques,while neglecting the direction of edges during domain adaptation.For network embedding in single-network scenario,this thesis presents a meta-learning based adaptive network embedding model called MLANE.The main idea of MLANE is capturing homophily and structural equivalence for different nodes in different tasks adaptively by incorporating sampling strategy learning with embedding learning into one optimization problem that can be solved via an end-to-end meta-learning framework.Experiments on realworld datasets verify the effectiveness of MLANE on various single-network analysis tasks.For network embedding in multi-network scenario,this thesis presents a domain adaptive network embedding for directed network alignment model called DADNA.By designing weighted adjacency matrices,DADNA can adjust weights of nodes with the direction of edges.Meanwhile,DADNA generates node embeddings using graph convolutional networks and achieves domain adaption between different directed networks by generative adversarial networks.Experiments on real-world datasets verify the effectiveness of DADNA on network alignment task,which belongs to multi-network analysis tasks.