Real And Complex Scenario-oriented Graph Convolutional Network For Network Embedding

Network Embedding which aims to learn a low-dimension vector representation for each node,can be used for many downstream tasks through vector-based machine learning methods,such as node classification,community detection,link prediction,anomaly detection.In recent years,with the development of deep learning,graph convolutional network(GCN)shows strong power in extracting effective node representation.GCN gradually replaces the traditional network representation methods,and is widely used in industries such as recommendation and search.However,existing GCN-based methods can not deal with attribute-missing networks(attributes of partial node are completely missing)and high heterophilic networks(most connected nodes belong to different categories),so that they can not be generalized to more complex and general application scenarios.This paper focuses on the two issues of GCN in network representation learning.The specific results are as follows:First,pointing at attribute-missing networks,this paper propose a graph convolutional network based on adaptive attribute completion,to realize network embedding for attribute-missing networks.This method mainly uses the generative adversarial network to model the relationship between real attributes and network structure,and generates more real attributes by mapping from hidden space to attribute space in order to fit the real attribute-structure semantic relationship,which ensures the authenticity and effectiveness of the generated attributes.At the same time,this method uses the principle of mutual information maximization to realize the mutual promotion and joint training of GCN-based representation learning and attribute completion process.This paper compares this method with eight other representative network embedding methods in a series of network analysis tasks to verify the superiority and robustness of this method,and demonstrate that this method can adaptively learn the representation of arbitrary attribute-missing networks.Second,pointing at networks with heterophily,this paper propose a novel graph convolutional network with an adaptive feature propagation mechanism,to realize network embedding for networks with either heterophily or homophily.This method introduces a learnable homophily degree matrix to model the homophily or heterophily of the network,so as to capture the potential homophily or heterophily distribution of the network.The homophily degree matrix is learned via extracting class-aware information from attribute space and structure space.Then,this method introduces the homophily degree matrix into the graph convolution framework to guide and control the propagation process,so as to achieve the purpose of adaptive learning.This paper compares this method with nine other network embedding methods,validating the superiority of this method,and demonstrating the adaptive learning ability of the this method.The research contents and methods proposed in this paper effectively overcome the limitations of traditional graph convolutional network in real and complex application scenarios,and expand the application scope of graph convolutional network.