Research And Application Of Network Representation Learning Algorithm For Clustering And Link Prediction

With the rapid growth of network data scale,traditional network representation methods have become a bottleneck restricting large-scale network processing and analysis.How to represent the original high-dimensional network data in low-dimensional vector space by network representation learning has become a problem worthy of research.An effective way is to map the network into a low-dimensional vector space,that is,to represent the network with low-dimensional vectors.Network representation learning can represent high-dimensional sparse network matrix with dense and continuous vectors in low-dimensional space.Network processed by the network representation learning algorithm will gets rid of the constraints of edges in adjacent matrix.Hence,each vertex in the network can be represented with a independent feature vector in low-dimensional space,which can thereby support subsequent applications.In this thesis,research on representation learning algorithm of static network and dynamic network is based on the local similarity and network dynamics of the network.The main content of research includes the following aspects:(1)In terms of the shortcomings of current static network representation learning algorithms,this thesis proposes a static network representation learning algorithm based on autoencoder.It first obtains the probabilistic co-occurrence matrix of the original network with random walks,then acquires the positive pointwise mutual information matrix of the network by combining the network vertices and the probabilistic co-occurrence matrix with word representation learning.Finally,it learns the structural information of the network with the semi-supervised autoencoder structure in deep learning,and the original network is thereby represented by a low-dimensional dense vector in the new vector space.The performance of this algorithm on different tasks is compared with algorithms that currently perform well on several metrics by performing experiments on different datasets.(2)Meanwhile,in view of the dynamics of real-world networks and the fact that there is still much room for development of dynamic network representation learning algorithms,this thesis proposes a dynamic network representation learning algorithm based on graph convolution.In order to process complex dynamic network data,it first learn the structure of the entire network with the idea of convolution,that is,process this non-European spatial data with graph convolutional neural networks.The network feature vector obtained is then processed with recurrent neural network to learn underlying information of time series in the dynamic network,thereby acquiring the characteristic vector of the dynamic network in a low-dimensional space.Finally,it reconstructs the representation vector in space with the decoding process to acquire the overall structure of the network at the next moment,and adjust the parameters in network representation vector according to the error between the original input data and the output data with back propagation.Hence,this algorithm can acquire network representation vector and predict the network structure of the next stage at the same time.(3)This thesis applies the algorithms proposed above to the expert collaborator network prediction system,so as to realize the prediction of expert cooperative relationship.The system predicts the relationship between experts with the network representation vector,which is acquired by training these algorithms on dataset DBLP and is then visualized with web application development technology.