Research On Key Technologies Of Representation Learning For Complex Networks

Recent years have witnessed the success of the network analysis.The network or graph is widely applied used to model various real-world applications,such as social networks,traffic network,world-wide-web,internet of things,gene network,etc..Network representation learning,also known as network embedding,aims to transform the components of the network(nodes,edges,subgraphs,etc.)into low-dimensional vectors,while maximally capturing the structure and characteristics of the original network.Because low-dimensional representation vectors can be easily processed by machine learning algorithms,network representation learning is receiving tremendous attention from the industry and academia.It is being successfully applied to a variety of tasks,such as node classification,link prediction,network reconstruction,recommendation system,etc.Researchers at home and abroad have made great efforts on network representation learning and accelerated its development.However,the existing works can not handle the global structure information,multi-typed attributes and potential relationships of the network effectively.In this thesis,we aim to address above issues and study the key technologies of representation learning for complex networks.The main contents and contributions of this thesis are described as follows.(1)The existing embedding methods mainly focus on capturing the microscopic structure of the nodes in the network.But they ignore the global roles played by the nodes,resulting in the limitations in the mesoscopic and macroscopic tasks.To solve this problem,we propose a network representation learning model named Rank2vec,which considers both local structure and global roles of the nodes in the network.The learned representation vectors can preserve both the microscopic and macroscopic information of the network.To evaluate the proposed model,we conduct some extensive experiments on the task of multi-label classification on several real data sets.The experimental results have shown that the Rank2vec achieves significant improvements than state-of-the-art methods.(2)In addition to the topological structure information,the real world network also contains abundant useful attributed information.At present,the existing studies on attributed networks cannot explore both the multi-typed attributes and the semantic relationships flexibly.To address the above problem,we propose a deep model based embedding learning method for attributed networks,named DeepEmLAN.Besides preserving the topology information,we design a deep attention model that smoothly projects different types of attributed information into the same semantic space.Furthermore,we design a heuristic combining strategy to generate the final embeddings,which makes the nodes sharing more neighbors,similar text-enriched or labeled attributes closer in the representational space.To evaluate the potential of the proposed DeepEmLAN,we evaluate its performance on the challenging tasks of node classification and network reconstruction.The experimental results on several real data sets have shown that DeepEmLAN outperforms competitive state-of-the-art methods significantly.(3)How to extend deep learning to graph-structured data has attracted a number of researchers.However,the existing works can not fully explore the potential relationships and structural characteristics between nodes,resulting in the poor robustness of network representation learning model.To solve the above problem,we propose a robust graph convolutional neural network model RoGCN by considering potential relationships and closed-loop structure.This model utilizes attribute information to mine potential relationships,and introduces graph convolutional neural network for inductive representation learning.In addition,a triangle-structure-based node selecting and aggregating method is designed in the proposed model.Based on the local structure of the target node,the method can aggregate more information of the neighborhood nodes with the same category,so as to learn the discriminative representations.Experimental results on the task of node classification show that the performance of the proposed RoGCN model performs better than the state-of-the-art algorithms.