Dynamic Network Representation Learning Based On Graph Convolution And Its Application

With the development of network science theory,networks have become a structure that can effectively describe the association between data,and network modeling is also widely used in various data analysis scenarios,becoming an efficient approach of modeling analysis.Using machine learning in the tasks of network modeling analysis is a popular and efficient way,and the effective network representation is the key to implement machine learning methods.At present,representation learning methods based on static networks have already been mature basically,however,more real data changes over time.Thus,data possessing the dynamic evolutionary characteristic can be described in networks language as the dynamic networks whose nodes and edges change over time.And effective representation learning for networks with a distinct dynamic evolutionary characteristic is the current focus of network representation.The aim of this thesis is to propose an effective method for learning dynamic network representations,mapping dynamic networks to low-dimensional vector representations,while being able to maintain the topological characteristics of the network as well as the evolutionary characteristics in terms of time.Therefore,this thesis designs and implements a new dynamic network representation learning model,which converts the network snapshot corresponding to each moment in dynamic network into a new target network node based on critical sub-graph similarities,and utilizes graph convolution to obtain a new target network representation that is a representation of original dynamic networks.The content of this thesis mainly includes two parts,the first part is the exploring study of critical nodes in the network,designing an effective method to find them.And the second part is on the basis of the first part utilizing the critical nodes of dynamic networks and the change degree of its sub-graphs’ structure to describe the evolutionary relevance of dynamic networks,then we reconstructs the dynamic network into a static“hyper-network”with evolutionary characteristic in according with the relevance.Finally,we build a framework of hyper-network nodes’ representation learning by graph convolution to obtain the low-dimensional vector representation of dynamic networks.The main innovations of the thesis are in the following two aspects:1、Proposing a fast and efficient algorithm for finding the critical nodes of networks;2、Proposing a method for describing network snapshot relevance in dynamic networks based on critical nodes and its sub-graphs similarities.In addition,we transform the time series corresponding to the dynamic network into a static network consisting of snapshots of the network based on the similarity of sub-graphs,and then transform the problem of dynamic network representations into the problem of new static network representations.In this thesis,supervised and unsupervised learning tasks are performed on multiple dynamic network datasets,and the results of the experiment demonstrate the effectiveness of the dynamic network representation learning framework proposed.Particularly,this thesis builds a dynamic network autoencoder framework for unsupervised dynamic network change-points detection and a dynamic network supervised learning framework for schizophrenia discrimination on fRMI data modeling.Compared to the current mainstream approaches,the dynamic network representation learning approach proposed in this thesis shows a significant improvement in algorithm performance on both unsupervised and supervised learning tasks.