Classification Graph Classification Based On Graph Convolution Networks And Explicit Tensor Representation Of Capsule Networks

Graph,as a kind of complex data structure,can represent many feature structures with rich information,and it can be used to describe various relationship between complex data features and the structural features.At present,graph structured data has been widely used in social networks,electronic shopping,chemical molecules and transportation networks.The classification task of graph structure data is to classify graph structure data according to graph nodes or structure features by learning the relationships among a set of labeled graph structure features without supervision,and to be able to correctly predict the invisible graph labels or structures.Because the graph structure data contains both node features and structure features,the complexity of feature information is too high,which increases the difficulty of data classification to a large extent.Moreover,due to the irregularity of the size of graph structure data,the application of Convolutional Neural Networks(CNNs)in graph data analysis is restricted.Mainly for large-scale different size figure structure data classification and data features respresentation,this paper studies the graph convolution network model based on global pooling and hierarchical pooling,and the capsule network learning model based on feature vector representation,in order to improve the classification accuracy of graph structure data.The research content of this paper is mainly from the following two aspects:(1)A graph classification method of graph convolutional network based on score-discarding pooling mechanism is proposedIn order to solve the problems of detail information loss and low classification accuracy in the feature extraction process of graph-structured data,this paper sets up a global pooling and hierarchical pooling graph convolutional network model based on score-discarding mechanism,starting from the difficult problems in the pooling operation of irregular graph-structured data.Firstly,the features of the graph data are combined and extracted through the graph convolution layer.The nodes are discarded by score in the graph pooling layer based on score discarding.Secondly,a higher level abstract feature map is formed through the stitching and collation of the feature data vectors after global pooling and hierarchical pooling.Then,the result features of the upper layer are classified through a fully connected network composed of multi-layer perceptron network and Softmax layer.Finally,experiments show that the algorithm model constructed in this paper performs better on four standard graph datasets,such as D&D,PROTEINS,NCI-1,and Frankenstein,than graph pooling network models such as Set2set,Sortpool,Diffpool,and Eigenpool.At the same time,it has a stronger ability to extract structural features.(2)A graph classification method based on capsule network representation of fixed size tensor is proposedIn order to better study the feature representation of extracting graph structure data and further improve the accuracy of graph structure data classification task.On the basis of the above research work,a graph classification model of capsule network based on the vectorization of graph data is constructed.Firstly,the large-scale graph data of different sizes are formed into 3D data blocks through the selection of node sequence,determination of the neighborhood range of root node and the process of graph standardization.Secondly,the 3D graph data blocks are processed into the form of capsule units through convolutional feature coding and input into the capsule network,and the node-edge data features are learned by using the dynamic routing algorithm between capsules.Then,in the process of feature learning,tensor extraction,feature representation in the convolutional layer and capsule network representation layer,T-SNE data visualization algorithm is used to map the high-dimensional data into the dimensional data space for visualization analysis.Finally,the proposed method was tested on three types of nine standard graph datasets,namely MUTAG,NCI-1,NCI-109,PTC,D&D,Proteins,ENZYMES,IMDB-B and IMDB-M,to verify the effectiveness of the proposed algorithm.