Research On Link Prediction Algorithm Based On Network Representation Learning

Link prediction is an important task for network analysis which also has a profound meaning in theoretical research and an important value in practical applications.For example,in the domains of theoretical research,it helps to understand the mechanism of network evolution;in practical applications,it can not only be used in various recommendation systems but also help biologists to study the interaction of proteins.Therefore,it has been concerned by scholars in many fields of science.Since the representation learning technology has achieved good results in solving problems of network analysis,more and more people begin to study the application of network representation learning algorithm in network analysis tasks.Link prediction is one of the most common tasks in network analysis.The existing random walk(RW)based network representation learning algorithms use RW or its variants algorithms to generate the sequence of nodes,but RW and its variants algorithms tend to choose nodes with bigger degree when traversing the network,which make the generated sequence of nodes can not reflect the network structure information very well and affect the performance of representation learning algorithms.To solve this problem,this thesis introduces unbiased sampling algorithm MHRW,and proposes a remove self-loop for MHRW algorithm(RLP-MHRW),then proposes a network representation learning algorithm based on improved random walk.In this thesis,the node representation vectors learned by the proposed algorithm are used as network features,and link prediction is realized by combining logistic regression model.In order to verify the effectiveness of the proposed algorithm,AUC and precision are used as evaluation indicators to simulate on four real network data sets.The experimental results show that compared with classical algorithms: CN,JAC,AA and network representation learning algorithms: DeepWalk,LINE and node2 vec,the performance of the proposed algorithm in link prediction is improved.In addition,this thesis discusses how to improve link prediction accuracy by using deep learning technology,and designs and implements a link prediction model based on residual attention mechanism(AM-ResNet-LP).The aim of AM-ResNet-LP model is to optimize the link prediction effect.The implementation steps of the model are mainly divided into three steps: firstly,data preprocessing,learning node representation vectors by using the network representation learning algorithm proposed in this thesis,using node representation vectors to assist in generating node's neighborhood subgraph information,and further merging the neighborhood subgraph information of a single node into the information data of node pairs and labeling them.Secondly,a classification model is built on the TensorFlow framework by using residual network and attention mechanism.Stochastic Gradient Descent(SGD)and Dropout technology are used to train and optimize the model.Finally,the trained optimal model is used for link prediction on real network datasets.The experimental result show that AM-ResNet-LP performs better than the previous link prediction algorithm based on representation learning in AUC index.