Research On Node Classification Model Based On Multi-level Graph Attention Convolutional Neural Network

In recent years,deep learning has developed rapidly and has achieved amazing results in many application fields.From speech recognition,machine translation to image recognition and computer vision,there are related theoretical breakthroughs and practical applications.However,with the development of information technology,the carrier carrying information is not limited to data such as text,audio and video or images.Relational data represented by social networks,citation networks,communication networks,and biological networks also contain a lot of information.Classification is an important way of data mining.The classification results can be used to obtain the interests,hobbies,values of the users in the social network or the research fields of the literature in the citation network.The problem studied in this paper is to complete the node classification task based on the deep learning method.The node itself contains a large amount of content information indicating the attributes of the node.The node and the node represent the connection information of the node through the edge.This paper hopes to extract the above two types of information and complete the single-label classification and multi-label classification task of the node. Graph Convolutional Networks(GCN)is a model that uses deep learning to implement node classification tasks.It can fuse the node's content information and connection information,and label the unobserved nodes.However,this model has the disadvantage of involving complex matrix operations and insufficient explanatory power.In view of the above problems,this paper has improved the model based on previous studies.The main research work is as follows: 1)This paper proposes a model for extracting node features by graph convolution and classifying nodes.The model accepts the topological graph data of the boundless weight matrix with auxiliary information on a node,and directly operates the nodes on the topological graph data to avoid Complex matrix operations,which integrate node auxiliary information and spatial structure information in a natural and intuitive way to obtain low-dimensional vector representation of nodes and implement further classification tasks.2)This paper introduces the attention mechanism so that the model can give different convolution parameters to each neighbor node in the convolution operation.This parameter weights and sums the neighbor nodes to realize the extraction of central node content information and local connection information.This model uses neural collaborative filtering(NCF)to achieve attention mechanisms.This model reduces the time complexity of the algorithm model by a more reasonable neighbor node extraction strategy,and emphasizes the retention of the central node feature information through a skip connection mechanism.3)The models and algorithms presented in this paper are based on the Python language and the Tensorflow deep learning framework to implement the algorithm model.The comparison experiments were carried out using two well-known benchmark datasets.The different algorithm models in the existing research fields were compared and verified, and the feasibility of the proposed algorithm model in node classification tasks was proved.In addition,the rationality of the model structure design is verified by design experiments.