Research On Graph Embedding Model Based On Deep Neural Networks

In the real world,many important data exist in the form of complex networks or graphs,such as citation networks,transportation networks,and genetic networks.There are a lot of value information in the characteristic information attached to nodes and the link relationship between nodes in the network.In addition,the space-time dependent information in the dynamic network is also of great significance for analyzing the changing trend of the dynamic network and predicting the behavior of the nodes in the network.However,the current graph embedding method does not implement network graph embedding well.For example,the matrix decomposition method can capture the structural information in the network,but it is limited by the huge amount of calculation and cannot deal with large-scale networks.Based on the random walk method,the structural information of the network can also be learned,but the global structural information of the network cannot be captured nor can the information such as node attributes and labels be used.For dynamic networks,the current graph embedding model mainly enhances the smoothness of the representation of the nodes in the adjacent dynamic network mirror by applying a temporal regularization,and assumes that the spatiotemporal evolution of the dynamic network lasts very short?such as assuming that only two steps of change?.These methods may fail when the nodes exhibit significantly different evolutionary behaviors.In addition,most of the current graph embedding models are based on shallow models,which are difficult to capture the deep features of complex networks.With the rapid development of deep learning technology,it has made great progress in many fields.This paper will focus on the graph embedding model based on deep learning,and according to the characteristics of static and dynamic networks,we propose a graph embedding model that can handle static and dynamic networks respectively.The specific research contents are as follows:?1?For static networks,in order to solve the problem that the current method cannot simultaneously encode network structure information and node feature information.This paper combines a graph convolutional neural network?GCN?and an autoencoder?AE?to propose a scalable semi-supervised depth graph embedding model—Semi-GCNAE.Use GCN to capture the structure and feature information of all nodes in the K-order neighborhood of node₄)in the original network,and use this as the input of AE.AE performs feature extraction and non-linear dimensionality reduction on the K-order neighborhood information captured by GCN,and combines the Laplacian feature map to preserve the cluster structure of the nodes.Through the introduction of integrated learning methods and joint training of GCN and AE,the model's learned low-dimensional vector representation can retain network structure information and node feature information at the same time.Extensive evaluations on five real data sets show that the low-dimensional vector representations of nodes learned by our model can effectively retain the structure and node feature information of the network,and compare with the existing tasks in node classification,visualization and network reconstruction Model performance has been significantly improved.?2?For dynamic networks,in order to solve the current dynamic network representation learning model,it is impossible to capture the high-order similar characteristics of nodes in dynamic networks.At the same time,it ensures that the model can effectively learn the spatiotemporal dependent information of dynamic network.In this paper,a dynamic graph embedding model—Dyn CNNLSTM is proposed based on the framework of autoencoder?AE?combined with convolutional neural network?CNN?and long and short-term memory neural network?LSTM?.It can simultaneously capture the higher-order structural features and spatiotemporal evolution features of the dynamic network,and generate a series of time series node vector representations for each node.The convolution operation in CNN layer is used to extract high-order similar features of nodes in the network.The LSTM layer is used to learn the time-space dependent information of dynamic network.In order to explore the influence of different length of historical spatiotemporal dependency information on the performance of the model,this paper also introduces the look back?lb?to control the length of spatiotemporal dependency of the model learning dynamic network.Through the link prediction experiments on one simulation data set and two real data sets,it is shown that the dynamic graph embedding model DynCNNLSTM proposed in this paper can effectively capture the high-order structure characteristics and network time-space evolution information of the dynamic network,which is significantly improved compared with the current model.