Research On Semi-supervised Node Classification Based Structural Network

Various complex network structures emerge in recent years with the rapid development of big data and artificial intelligence fields,like social networks,paper citation networks,and genetic engineering networks.These networks contain a large amount of valuable information,and attract many scholars to conduct network research to explore important information in the network.For example,social network explores the relationship of users to obtain the users' relevance information,which input into the classification model,so that discovering the users' interest preferences to better serve the fields of advertisement injecting,marketing,community communication.The reality is that supervised learning tends to be harder owing to a mass of unlabeled nodes in the network.The early solution is to manually label nodes by expert experience.Obviously this method is a big expense either time or finance.It's a hot issue of how to use the existing semi-supervised classification learning model to classify and predict unlabeled nodes in the network.Different from the general classification problem,nodes in the network not only have their own node attributes,but also network information based node relationship.Till now there are two mainstream node classification methods: node classification based on representation learning and node classification based on deep learning.The former only considers information from the network structure to generate a vector representation of the node,which applies to the node classification task.The latter implements node classification based on the network structure and the features of the nodes themselves.After analyzing many node classfication methods based presentation learning,it is clear that they ignore the difference in the network structure of different nodes like the degree of the node.This paper researchs the existing representation learning algorithms and semi-supervised node classification algorithms,and proposes a semi-supervised classification algorithm based on network structure information.The improvement shows below:(1)A penalty factor is set for each node of the network to limit the length of the node's random walk sequence so that different nodes have different lengths of walk sequences,which are equivalent to sentences in natural language processing.The node sequences are as input to the word2 vec model for the purpose of translating the potential information of the network structure into vectors.(2)We improve the semi-supervised classification algorithm based on deep neural network.Pseudo labels are generated for unlabeled nodes,and then mixing labeled nodes and unlabeled nodes,next MixUp method is used for data augmentation to obtain a new training set.The classification loss of the unlabeled nodes is added to the loss function of the model to achieve optimal for the model.Finally,we conduct a series of experiments comparing with the four famous methods on the three standard datasets.The results show that the proposed method has better classification effect than the existing methods.