Research On Semi-supervised Graph Neural Network Method Based On Graph Fusion

Machine learning methods have been used in all aspects of production and life.However,the accuracy of unsupervised machine learning method can not meet the actual production demand in complex data.The accuracy of supervised learning method is high,but it is heavily dependent on a large number of marked training data,and the cost of manual marking data is expensive.Therefore,semi-supervised learning methods which can be used to learn jointly with a small number of mark samples and a large number of unlabeled samples,have been favored by industry and academia.However,the current graph based semi-supervised learning methods,such as graph neural network,rely heavily on a single fixed graph,which restricts the performance of graph based semi-supervised learning methods.This paper presents a neural network method for constructing multi graphs dynamically,aiming at alleviating the dependence of graph based semi-supervised learning method on a single fixed graph and improving the performance of the semi-supervised learning method based on graph.In this paper,a compact graph fusion convolution framework is proposed.This framework can be divided into three parts,including the part of constructing the similarity matrix of multi graphs,the expanding part of the similarity matrix of multiple graphs,and the fusion part of the similar matrix of multiple graphs after the expansion.In the part of constructing the similarity matrix of multi graphs,two methods are proposed,including Markov chain method based and multi metric method.The similarity matrix of multiple graphs obtained by dynamic construction is obtained from different angles,which can better reflect the similarity between samples from many aspects.For these two different methods of constructing the similarity matrix,different graph constraints are introduced in this paper,which forces the same class of samples to form more compact clusters in the hidden space by feature extraction network.In the graph enlargement part,we propose two kinds of operations to expand the original data similarity relations to multiple dimensions.We can use this multiple dimensions information to describes the similarity between any two samples.The graph augmentation operation can extract more information from different graphs.we also state that the graphs obtained by the augmented operation have changed from the original unstructured data to semi-structured data,which makes it possible to use traditional convolutional network on the expanded graph data.In the graph fusion part,in order to integrate the information which shows the similarity of samples in different dimensions,we use convolutional network to fuse the semistructured graph data obtained by enlargement operation to get better similarity matrix.The expanding and merging the multiple graphs method proposed in this paper,is a new space-based convolutional neural network method.Finally,the experiment shows that the method can get better graph similarity matrix and improve the accuracy of graph based semi-supervised learning method.The method proposed in this paper integrates similar information from multiple perspectives,combines the advantages of graph regularization and self expression constraints,and can be easily applied to existing networks and use the information on unmarked data effectively.