Research On Dynamic Network Link Prediction Technology Based On Deep Learning

Dynamic networks whose network structure changes over time exist widely in reality.Links,as an important part of dynamic network structures,contain a wealth of information.Dynamic network link prediction can promote the research of network evolution mechanisms and the development of network science,it can also be applied to the fields of recommendation systems,biological experiments and so on.However,most previous researches focused on static networks,many link prediction methods only consider the structure of the network and ignore the information about the time evolution of the network,leading to accuracy limitations.Therefore,how to combine the structure information of the network and the temporal evolution information to improve the accuracy of link prediction in dynamic networks has become a research hotspot;In addition,links in the network may contain weight attributes,and these weight information often have physical meaning.The weight prediction of the links can describe the network more accurately.However,it is easy to cause structural errors in weight prediction.How to balance the accuracy of weights and the accuracy of structure has become a challenging problem in the current research.This paper is mainly based on deep learning technology to study the problem of dynamic network link prediction.For undirected and unweighted dynamic networks,a dynamic network link prediction model based on the self-attention mechanism is proposed.Futher more,for undirected weighted dynamic networks,a dynamic network link prediction model based on generative adversarial network is proposed;The effectiveness of the proposed model is verified by comparative experiments.The specific work is listed.Aiming at the dynamics of undirected and unweighted dynamic networks,the dynamic network is described in the form of graph snapshots,so the link prediction problem of the dynamic network can be approximated as a time series analysis problem.Based on the current progress of deep learning in sequential data learning,the self-attention mechanism is used to learn the temporal evolution information of the network,and the autoencoder is used to learn the structural information of the network,thus forming an end-to-end dynamic network link prediction model AESAN(Autoencoder and Self-Attention Network).Comparative experiments are carried out on three realistic dynamic network data sets:Contact,Enron and Ant,and compared with the link prediction algorithm based on topological structure and exponential average weighting and other deep learning models.The experiment results show that the AESAN model proposed in this paper has certain advantages over the algorithm based on topology in the accuracy metrics.Although it is not as good as the deep learning model using long and short-term memory networks in AUC,Precision and other metrics,it has certain advantages in the training time.That is,relatively accurate prediction results can be obtained in a short time.Aiming at the problem of link weight prediction of undirected weighted dynamic networks,inspired by the excellent performance of generative adversarial network structure in sequence data prediction and graph convolution network in graph representation learning,a dynamic network link weight prediction model based on generative adversarial network Att-GAN(Attention based Generative Adversarial Network)is proposed.The model uses graph convolution network and self-attention mechanism to construct a generator,uses fully connected layers to form a discriminator,and optimizes the model by the adversarial training to get high-quality prediction results.Comparative experiments are carried out on three realistic weighted network data sets:UCSB,Enron and Cell-calls.Compared with link prediction algorithms based on synthetic networks,link prediction algorithms based on matrix factorization,and other deep learning models,AttGAN has certain advantages in the index of measuring the weight error,and performs well in the index of measuring the structural error,which means the model can generate high-quality,robust prediction results.