Research On Attributed Network Representation Learning Methods Based On Deep Learning

Network representation learning,also known as network embedding,aims at representing nodes in a network as low-dimensional and dense real-valued vectors.Network representation learning provides an efficient way to represent a network and the learned representations can be flexibly and conveniently applied to various network analysis tasks such as node classification,link prediction,community detection,and so on.In real life,nodes in networks are often associated with descriptive attributes and these networks are named attributed networks.In recent years,attributed network representation learning based on deep learning has gradually attracted a surge of attentions from researchers.However,the lack of effective use of community structure and task-related information is still a problem in most existing attributed network representation learning methods.Aiming at solving the problem,this thesis proposes two effective attributed network representation learning(network embedding)methods.The main contributions are as follows:(1)To solve the problem that community information cannot be used effectively,a multi-view features preserved network embedding method is proposed.In this method,the features of nodes are constructed from the views of local network structure,nodes attributes,and community structure,respectively,and then the nodes representations are obtained by integrating the results of multiple autoencoders so that the nodes representations can reflect a variety of similarities among nodes in the network.Besides,a topological constraint and a consistency constraint are introduced to enhance the robustness and consistency of the learned nodes representations.(2)To solve the problem that task-related information cannot be effectively used,a task-oriented attributed network embedding method is proposed.This method extracts various features based on attributes of the different neighborhoods of nodes and fuses them to obtain nodes representations.They are then inputted into three functional modules to preserve the similarity between nodes in terms of network structure and nodes attributes,as well as task-related information,respectively.Because the task-related information is incorporated into the learning process,the method can be flexibly used for different network analysis tasks and make the learned nodes representations match the goal of different tasks.The network representation learning methods proposed in this thesis are fully tested by experiments on well-known and widely used network datasets.Experimental results show that the performances of the proposed methods are better than the state-of-the-art methods,which verifies the effectiveness of the proposed methods.