Research On Deep Graph Convolutional Network Model Algorithms With Local And Global Consistency

With the rapid development of the computer network technology and the diversification of data acquisition methods,the information available to people is growing exponentially.How to effectively utilize these massive data to accelerate the development of social productive force is one of the common challenges for global researchers and technical experts at present.In recent years,machine learning and deep learning have developed rapidly.It is well known that training a deep neural model typically requires a large amount of labeled data,which is hard to get in many scenarios due to the high cost of labeling training data.Semi-supervised learning(SSL)is an important sub-field in machine learning and plays a key role in modern intelligent technology.Graph-based semi-supervised learning(GSSL)is the mainstream method in the SSL.GSSL mainly uses graphs to describe the space structure of data.Owing to the powerful representation capability of the graph data,many real-life scenarios,such as transportation networks,social networks,and citation networks,are located in the form of graphs.In these scenarios,graph-based semi-supervised node classification has attracted much attention due to the wide range of applications,e.g.,user tagging in social networks and product recommendation.Both graph convolutional network(GCN)and local-global consensus(LGC)algorithm are message passing algorithms,which have achieved superior performance in semi-supervised classification.GCN is typically based on stacked message-passing layers that share neighborhood information to transform node features into predictive embeddings.In contrast,LGC involves spreading label information to unlabeled nodes via a parameter-free diffusion process,but operates independently of the node features.By analyzing the shortcomings of some existing algorithms,this thesis,based on previous studies,improves the learning model.The main research contents are as follows:GCN follows the propagation mechanism similar to LGC.The main difference is whether features or labels are smooth across the graph.Then,it is natural to consider a combination of the two for using the feature and label information of the graph data to improve the classification performance of the model.In this thesis,we combine GCN and a barebones LGC(BLGC)algorithm with few parameters for node classification to create a unified model BLGC-GCN.In this model,BLGC as a regularizer can increase the capability of GCN of simultaneously learning the transformation matrix and the edge optimal weight matrix.The optimal edge weight matrix maximizes the probability of each node being correctly marked by BLGC,and also increases the influence of intra-class marks and intra-class features of each node,which is conducive to the separation of nodes of different classes.The optimal edge weight matrix and corresponding degree matrix learned by BLGC are then applied to a GCN model to predict node labels.Extensive experiments on five benchmark datasets have been performed,and the experimental results show that the classification performance of BLGC-GCN is significantly better than the state-of-the-art baseline methods.This demonstrates the effectiveness of the proposed BLGC-GCN model.GCN provides a completely new way of learning graph data and has achieved excellent performance in semi-supervised learning performance.However,GCN model cannot be stacked as deeply as the CNN model in visual tasks.Once multi-layer GCN is used,the classification effect of related tasks will decline sharply,which seriously limits the capability of GCN in some tasks.The main reason for this is excessive smoothness,that is,after multi-layer GCN is used,the differentiation of nodes of different categories becomes worse and worse,and the representation vectors of nodes tend to be consistent.In this thesis,we further study the over-smoothing problem in deep models,and propose a barebones Layer normalization(B-Layernorm)method with few parameters for GCN,referring to the layer normalization method in natural language processing.B-Layernorm has the effect of preventing the output characteristics of distant nodes from becoming extremely similar or indistinguishable,and meanwhile allowing the output characteristics of connected nodes in the same cluster to become more similar.B-Layernorm is generalized to the unified model BLGC-GCN and a new model BLGC-GCN*is obtained.Meanwhile,residual connections are added for each layer during model training to avoid the training difficulties caused by gradient disappearance.A large number of experiments have been carried out on three reference network data sets,and the experimental results verify that B-Layernorm can improve and alleviate the over-smoothing problem in the deep model,making the model more robust to over-smoothing.Compared with the latest baseline method,the experimental results show that BLGC-GCN*has obvious advantages and can make the deep model get the same results as the shallow models.