Research Of Deep Network Embedding Method Based On Generative Adversarial Network

Nowadays,as a widely used data carrier,network is gradually becoming one of the ways that people recognize and abstract the world.In addition to the node and edge information,the network often includes rich node attributes,which has great value of information mining.Network representation learning,is a novel representation learning method designed to represent nodes in network as a low-dimensional,dense,and real-valued vector representation.The learnt vector representations will be used in various downstream tasks such as node classification and link prediction to improve the performance.However,the existing deep network representation learning methods often fall into the overfitting problem because they ignore the distribution of the learnt representation,which means that this will affect the generalization ability of the learnt embedding vector representation in downstream tasks.Generative adversarial network uses the idea of adversarial training,by setting a generator to fit the true distribution of the data.Network representation learning methods which based on generative adversarial network,use the adversarial training mechanism to constrain the distribution of the learnt embedding representation to alleviate its overfitting problem.However,the existing methods are mainly unsupervised learning approaches,and there are two problems:first,the performance of extracted network embedding is insufficient,and the neighboring information of the network is not fully utilized,and they also lack of reconstruction of the content information;second,the adversarial training mechanism used by these methods need to specify a prior distribution manually,and there is a problem of mode collapse,at the same time,the classic adversarial loss function has the difficulty in generator training.In view of the above problems,this thesis proposes an attention-based adversarially regularized network representation learning model and a Wasserstein distance based adversarially regularized network representation learning model.The main work includes:(1)To address the poor performance of extracted network embedding,an attention based adversarially regularized network embedding model AARNE is proposed.This model is inspired by GAT on the basis of the ARGA model.An attention autoencoder is proposed to learn the embedding representation of the network.The attention mechanism is used to dynamically consider the weights of different neighboring nodes when aggregating the neighbors' information to update the node representation.At the same time,the adjacency matrix and the content matrix are reconstructed by the structure decoder and the content decoder respectively.In this way,the embedding representation can simultaneously retain the structural information and content information of the network.AARNE also uses adversarial training mechanism to constrain the distribution of the learnt embedding representations to ease the overfitting problem.Finally,experiments on Cora,Citeseer and Pubmed datasets verify that the AARNE model can learn better embedding representations.(2)Further,to address the problems of mode collapse caused by manually specify a prior distribution in adversarial training and the problem of gradient vanish/instable caused by the usage of the traditional GAN loss,a Wasserstein adversarially regularized network embedding model WARNE is proposed.This model is inspired by WGAN,by introducing a generator to parameterize the prior distribution,the model collapse problem caused by manually specify a prior distribution is circumvented,and the loss function of adversarial training is modified based on Wasserstein distance to alleviate the phenomenon of gradient vanish/instable.The difference between the distribution of learnt embedding representation and the distribution of real data is minimized by reduce the new adversarial loss,and the embedding representation is regularized during the training process.Finally,experiments on Cora,Citeseer and Pubmed datasets verify the effectiveness of the WARNE model.