Research On Presentation Learning Methods Of Semi-supervised Network Based On Deep Learning

In semi-supervised network representation learning,node labels play an important role in guiding the establishment of network mapping relationships in different spaces.However,in many practical tasks,the available labels are limited,difficult to obtain or unevenly distributed.As a result,the learned network representations are not distinguishable enough in the downstream tasks such as node classification and node clustering,and it is difficult to ensure the validity of the representation results.Therefore,this paper carries out some targeted research,the details are as follows:(1)In the semi-supervised network representation learning,due to the limited available label information,the learned network representation is not sufficiently differentiated when performing downstream tasks.In this paper,a dual-channel semisupervised network representation learning model is designed,which is based on the autoencoder framework and consists of two information transfer channels: self-supervised and semi-supervised.The self-supervised signal and label information provide guidance for the establishment of network representation mapping relationship in two channels respectively,and form information complementarity and enhancement between them.Considering the possible information redundancy between two channels,a mechanism of redundancy recognition and elimination is designed in the view of mutual information.On this basis,an integrated optimization model is constructed to realize the synergistic effect of self-supervised learning and semi-supervised learning,so that the learned network representation can better capture and maintain the structure and characteristics of the network.Experimental results on real datasets show that the network representations learned by the proposed model can achieve better performance than baseline methods in tasks such as node classification,clustering and visualization.(2)In the semi-supervised network presentation learning,the distribution of label information is unbalanced,that is,the label information of each category is not equal,and some classes have no label information.This paper designs a semi-supervised network representation learning model based on completely unbalanced labels.The model is mainly composed of graph auto-encoder module,self-training clustering module and class semantic information module.Graph auto-encoder module is used to generate network representation while ensuring that the generated network representation keeps the original network information;Self-training clustering module and class semantic information module can capture and learn unknown class information.The self-training clustering module makes the nodes of the same class as close as possible and the nodes of different classes as far away as possible by dividing the distribution of node categories.The class semantic information module uses the known label information to generate class semantic information to guide the training of the supervision model,making the learned network representation more discriminative.Compared with some typical semi-supervised network representation learning algorithms on real data sets,the experimental results of this model are significantly improved.(3)A semi-supervised network embedded system is designed and implemented.The system integrates some classical algorithms and two semi-supervised network representation learning algorithms proposed in this paper.It mainly includes three modules: data loading,initialization and model evaluation.It can download the network embedding vector learned and carry out tasks such as node classification,clustering and visualization.