Research And Application Of Autoencoder-based Network Representation Learning Method

Networks in the real world are ubiquitous,how to efficiently analyze and process more and more large networks is very important nowadays.A key problem exists in the research of network analysis,how to reasonably represent the characteristic information of the network.With the development of machine learning,the characterization of network node information has become an emerging research field.Network representation learning aims to extract and embed the features of nodes or edges into a low-dimensional euclidean space to get a dense representation of the network,and then use some machine learning algorithms to perform various applications,such as link prediction,node classification,and community detection for the network,etc.This paper studies network representation learning,proposes two new network representation learning methods,and uses the network representation obtained by network representation learning to optimize the robustness of the network.The main research contents of this article are as follows:(1)proposes a deep stacked autoencoder-based network representation learning method(Encoder-AEO).First,the initial representation of the network is obtained through random walk with restarts,and then it is transformed into the positive pointwise mutual information matrix.Then,the feature information of the representation matrix is encoded by a deep stacked autoencoder.A new mixed objective function which can reflect the local and global importance of the network nodes is designed to achieve a more accurate low-dimensional representation for Encoder-AEO.On eight real network datasets,the new algorithm is compared with four well-known network representation learning methods in different tasks including link prediction,multi-label classification and visualization.The results show that the proposed new method has better performance.(2)proposes a network representation learning method with sequence autoencoder based on recurrent neural network.A long and short term memory network is used as a part of the hidden layer of the autoencoder to construct a sequence autoencoder.The model reconstructs the adjacency matrix of the network to extracts the global structure features.At the same time,the neighborhood information of the network nodes is encoded into the long and short term memory network by the sequence autoencoder,which effectively maintains the information features of the local structure of the network.And then the model performs deep training to obtain low-dimensional and dense representation vectors of network nodes which have undergone time series processing.The embedded representations of network nodes are applied to tasks such as network reconstruction,link prediction,and multi-label classification.The excellent results show that the proposed sequence autoencoder has a good network representation effect.(3)As the scale of the network becomes larger and the topology becomes more complex,it becomes more difficult to measure and optimize robustness.In order to reduce the time cost of robustness computation,this paper proposes a network robustness optimization model(LSNE-RO)based on the iterative local search algorithm and network representation learning method.First,a new attack form is proposed for the network representation with embedded,the attacks of nodes and connected edges are converted into the attacks on node embedding vectors and the embedded similarity matrix.Then the network structure is adjusted without changing the network degree distribution.Combined with the iterative local search algorithm,the network representation is analyzed to get robustness,and the optimal network structure is obtained during continuous iteration.Experiments on robustness optimization of four networks have found that LSNE-RO significantly improves the robustness of the network to malicious attacks on nodes,especially the LSNE-RO(DNGR)model combined with deep autoencoders improves the robustness of the network without increasing the time cost.