Research On Semi-supervised Graph Classification With Graph Convolutional Network

As a ubiquitous data structure,graph is widely used in social networks,recommendation systems,biochemistry and financial systems due to its powerful representation ability.Because these graph data commonly stores a lot of valuable information,many scholars study graph data to explore the important information in the data,and node classification is a common task in graph data research.For example,in the protein molecular network,classification learning is performed by analyzing the relationship between the protein network,so as to obtain the relevant properties of the protein.It is not only conducive to understanding the properties of proteins,but also has guiding significance in the fields of pharmaceuticals,agriculture and forestry science and technology.However,nowadays data is growing at a geometric speed,and there are a large number of unlabeled nodes in the data,and it is difficult to use supervised learning.In the early days,the prior knowledge of experts was used for manual data labeling,but labeling a large amount of data was costly and time-consuming and laborious,so scholars proposed semi-supervised learning.Different from supervised learning and unsupervised learning,semi-supervised learning solves the problem of the lack of large amounts of labeled data on the basis of ensuring the good generalization ability of the model,and at the same time improves the efficiency of manual labeling and the performance of model learning.How to use the existing semi-supervised classification learning model to classify and predict the unlabeled nodes in the graph network has become a hot topic.With the rapid development of deep learning,graph convolutional network(GCN)have become a powerful tool for processing irregularly structured data on graphs,and have achieved satisfactory results in graph representation learning tasks such as node classification.This paper analyzes the existing GCN algorithms and semi-supervised learning algorithms,and finds that the graph structure is used to guide the graph convolution operation in semi-supervised GCN.When the graph structure is inaccurate or even unavailable,the graph structure can be inferred or learned from the data to guide subsequent convolution operations.However,existing researches often construct graph structures based on simple distance metrics(such as k-nearest neighbor graph).Simple graph structures may not be able to fully mine the similar relationships between nodes,resulting in poor node classification task performance.Therefore,this paper will address the issues existing in the graph structure of the existing GCN models.Based on adaptive graph learning,low-rank learning,sparse learning and multi-graph learning,two methods for generating high-quality graph structures are proposed and used in GCN for semi-supervised classification.The main contents are as follows:(1)Semi-supervised learning with graph convolutional network based on hypergraph.Most of the existing GCN ignores the quality of the graph structure,which leads to unsatisfactory classification performance.To solve this problem,we propose a new graph learning method to output a high-quality graph structure,aiming at eventually improving classification performance for the downstream GCN model(HS-GCN)in this paper.Firstly,the proposed graph learning method employs an adaptive graph learning to capture the intrinsic low-level correlation of data,and learns the more useful high-level correlation from a hypergraph.Secondly,sparse learning and a low-rank constraint are integrated with graph learning respectively to remove redundant information.Finally,a compact graph structure is obtained to promote the information aggregation of GCN.The experimental results show that the graph structure of our proposed graph learning method can significantly enhance the classification performance of GCN.(2)Graph convolutional network based on multi-graph learning for semi-supervised classification.In general,the obtained data graph in graph convolutional network is a single graph.However,GCN only uses a single graph to guide graph convolution operations,which does not make use of the relevant information between various multi-graph.To solve this problem,we propose a graph convolutional network based on multi-graph learning for semi-supervised classification in this paper.First,multi-graph learning combined with a low-rank constraint is used to reduce the impact of redundant information.Then a shared high-quality graph structure is learned from multi-graph data,and we input the graph structure to graph convolutional network for semi-supervised classification.The experimental results show that the multi-graph learning method proposed is superior to other comparison methods in the semi-supervised classification task of GCN.Aiming at the problem that most of the existing GCN ignore the qualities of graph structures,this paper improves the graph learning method and uses the semi-supervised classification algorithm as an experimental evaluation method.The experimental results show that the proposed graph learning methods are better than the selected comparison methods in the performance of semi-supervised classification of GCN.In the future,I will consider how to learn graph dynamically,and integrate graph convolution into a unified framework,so as to obtain a graph structure that is more suitable for GCN.I will also explore the influence of different multi-graph fusion methods on graph structures in semi-supervised GCN models.