Research On Network Representation Learning Method Based On Adversarial Graph Convolution

With the rapid development of Internet,the network data is everywhere in our daily life,how to excavate the information in the network it is very important for us to use.Network representation learning is a common data mining method.However,the existing network representation learning methods ignore many key problems,such as ignoring the data distribution of hidden variables and failing to make full use of the real network data.Based on this,a novel network representation learning framework AGCN(Adversarial Graph Convolutional Networks)was proposed to solve the above problems.At the same time,an innovative end-to-end multi-task Learning model,MTL(Multi-task Learning),is proposed based on AGCN,considering the correlation between different tasks in the process of network representational learning.The main research contents of this paper are as follows:1.We proposed the AGCN framework.Based on the Autoencoder,we introduce GCN network and GAN network,and combine them to form AGCN model.The application of GCN network can not only capture the topological structure of the network,but also make full use of the attribute information of network data,so as to improve the accuracy of the model.For the feature vectors generated from the GCN network,GAN network is applied to force the feature vectors to match the prior distribution.In this way,the effect of network representation learning will be further improved.2.We proposed the MTL framework.This framework is an end-to-end multi-task learning framework.We fully consider the connection between node classification and link prediction,so that the two tasks can be trained and learned in a unified framework and two operations can be performed simultaneously in a single learning stage,thus improving the efficiency of the model.In MTL framework,based on AGCN and extended on the basis of AGCN,a two-layer MLP is used in the decoder layer of the Autoencoder to obtain the feature vectors.The feature vectors obtained from the encoder are transported to the first layer MLP of the decoder for node classification,and then the vectors obtained from this layer are transported to the second layer MLP for link prediction.In this way,they complement each other and constitute MTL framework.Experimental results on three famous citation network data sets also prove the effectiveness of the proposed AGCN model and MTL model.