Research On Graph Classification Method Based On Capsule Graph Neural Networ

In recent years,with the successful application of convolutional neural networks in various fields,people have been able to model and mine relevant information on regular Euclidean data.At the same time,people’s attention gradually shifted from regular Euclidean data to more complex non-Euclidean data,and proposed graph convolutional neural networks based on convolutional neural networks,trying to use neural networks to model graph data with more powerful representational capabilities.Many data in real applications can be modeled by graph data,such as biochemical molecules,social networks,knowledge graphs,etc.Therefore,mining knowledge from graph data has very important theoretical and practical significance,among which graph classification is an important branch in the field of graph mining,which is of guidance for drug discovery and network analysis.The main research of this paper is as follows.(1)In this paper,we propose a capsule graph neural network with EM routing(Caps GNNEM).Firstly,the primary capsule is constructed by extracting the features of nodes and topological information of the neighbor through graph convolutional neural network,then the capsule is updated by EM routing algorithm replacing the graph pooling layer to get higher level feature representation,and finally the whole graph level class capsule is obtained for graph classification task.(2)This paper also proposes a disentangle capsule graph neural network(DCaps GNN),the idea of which originates from the reason that the connection of graph data is heterogeneous in real life,so in order to better restore the original information,we first disentangle the features of the nodes by a feature decomposition algorithm to decompose the heterogeneous features,and then use the decomposed features for the capsule graph neural network,and finally obtain feature representations on the graph hierarchy for the graph classification task.(3)The two proposed graph capsule neural network models are encapsulated,and a molecular prediction system based on capsule graph neural networks is designed and implemented.The system provides complete functions for parameter setting,model training and model testing,and implements a simple interface to interact with the user to enhance the user experience.