Covariance-based Adversarial Autoencoder For Graph Embedding

This paper mainly studies the structure analysis and expression learning of large-scale graph network data.At present,various types of large-scale graph network structures are very common in real life,such as social networks,citation networks,transportation networks,Internet,logistics networks,biomolecular networks,etc.Based on these network structure data,there are many graph analysis tasks that need to be solved,such as link prediction(such as friend recommendation on social networks),multi-label classification(such as the classification of papers on the same topic on the citation network),and abnormal point detection(such as traffic Online congestion point detection)and so on.However,the traditional graph theory or statistical methods faced this large-scale and complex network new task,which was slightly inefficient and low performance.In recent years,researchers have proposed the use of machine learning to map graph network structures to low-dimensional vector spaces,and then classify and link predictions on graph structures through classification and similarity prediction in vector spaces.this method is collectively called graph embedding(or graph representation learning,network representation learning).In order to be able to better retain various node information and edge link information or other hidden information in the original network structure during the graph embedding process,a lot of work has done a lot of research on the graph embedding algorithm.In this paper,we mainly carried out research on the embedding algorithm of homogenous graphs,and verified the effectiveness of the proposed algorithm through comparative experiments.Specific research contents include:(1)In view of the problem that some node attribute information in the current algorithm has not been considered,we propose a new variational autoencoder model based on covariance,and construct a graph convolutional neural network and our algorithm based on covariance variational autoencoders.We verified that this algorithm not only has the original advantages of graph convolutional neural networks,but also fully considers the correlation between node attribute information.Therefore,this algorithm has the advantages of more fusion and comprehensiveness,and is more suitable for some graph which have feature information on certain graph analysis task.(2)On the basis of the first algorithm proposed in this paper,in order to make our algorithm suitable for larger-scale graph embedding,and have higher stability androbustness,we propose a variational autoencoder based on covariance adversarial graph embedding algorithm.Afterwards,we verified its advantages and stability in practical tasks through relevant experiments.(3)A series of comparative studies with other algorithms are conducted for the two algorithms previously designed.We apply them to large-scale graph network datasets in reality,and apply the generated low-dimensional vectors to various realistic tasks.The graph analysis tasks included in this paper include link prediction,multi-label classification,visualization,etc.The experimental results show that our proposed method can be better than the previous algorithms in the field to some extent,verifying that the new algorithm proposed in this paper has certain Effectiveness and wide practicality.