Network Representation Learning Based On Auto-Encoder And Hyperbolic Geometry

Network representation learning,also known as network embedding and graph representation learning,aims to retain the required features of the network,represent the network as a node embedding set in the potential space,obtain the low-dimensional dense vector representation of the network,and then apply the vector representation of learning to network analysis tasks such as node classification,link prediction,node clustering,etc.Network representation learning can efficiently analyze and process network structure data,so the research on network representation learning has great significance and value.With the rapid development of the era of big data,network data is growing exponentially,resulting in an increasingly large network scale and more complex structure.However,there are still many challenges in the research of online representation learning.On the one hand,the network contains rich node attribute information,and the existing network representation learning methods consider only the network structure information and attribute information.On the other hand,the network often exhibit potential geometric hierarchies,while most representation learning methods only learn in euclidean space,which cannot effectively capture the potential geometric hierarchies of the network,resulting in representation distortion and dimension explosion.In view of the above problems,this thesis proposes the attribute representation learning method of dual-channel auto-encoder and the representation learning method of hyperbolic geometric transformation,to learn the low dimensional vector representation of network nodes.The main research contents and contributions are summarized as follows:Learning method of attribute network representation based on dual auto-encoder is proposed.Aiming at the single consideration of network topology information and attribute information,a dual channel auto-encoder model is designed as follows: structural auto-encoder one channel,based on multi hop attention mechanism,using the normalized adjacency matrix and characteristics of the network to expand the exploration of neighbor information and capture global and local structural information of the network.A low-pass Laplacian smoothing filter is designed to process attribute information along the path of the attribute auto-encoder,combining with the network adjacency matrix to aggregate similar node information to the target node.Finally,the vector representations obtained from these two self-encoders are adaptively fused to realize the interactive fusion learning of network structure information and attribute information.The experiments of node clustering and link prediction are carried out on four public data sets,and the comparison is made from the aspects of normalized mutual information and running time.The experimental results show that the proposed representation method has advantages.In view of the lack of information about the underlying geometric hierarchy of networks in traditional representation learning methods,a hyperbolic geometric network representation learning method is proposed to capture more comprehensive structural information of networks.Firstly,the geometric interaction between the two spaces is realized by transforming the euclidean space and the hyperbolic space through exponential and logarithmic maps.Secondly,an attention method based on hyperbolic distance is proposed,which applies attention mechanisms to hyperbolic embedding to learn neighbor node information in hyperbolic space,capture the geometric hierarchy of the network,and retain node attribute information.Finally,the hyperbolic representation vector is fused with the traditional euclidean representation vector to obtain a node representation vector that ultimately retains more structure and attribute information of the network.On the public dataset of the citation network,experiments on node classification and link prediction were conducted,compared with the benchmark algorithm,and ablation analysis was performed.Experimental results show that the method proposed in this thesis can improve classification accuracy,and verify that the proposed method has good performance and certain effectiveness.