| In recent years,network representation learning has attracted more and more attention as an effective method to analyze heterogeneous information networks by representing nodes in a low-dimensional space.Random walk-based methods and metapath-based methods are currently commonly used methods for network representation learning.However,these methods are mostly based on shallow neural networks,and it is difficult to capture heterogeneous information network structure information.Graph convolutional neural network(GCN)is a popular method for deep learning of graphs,which can make better use of the network topology,but the current GCN design is directed at a homogeneous information network,ignoring the richness of the network semantic information.In order to effectively mine the semantic information and highly non-linear network structure information in the heterogeneous information network,and thereby improve the effect of network representation,a graph convolutional of fusion meta-path based heterogeneous network representation learning algorithm MG2 vec is proposed in this paper.The main research contents are as follows:(1)To solve the problem that graph convolutional networks ignore the rich semantic information in heterogeneous information networks,the meta-paths in heterogeneous information networks are first processed,and then the relatedness measurement of the linked target objects is carried out based on different meta-paths,and the association matrix containing semantic information is constructed by calculating the relatedness among all the target objects.(2)To solve the problem that the shallow model cannot effectively capture the heterogeneous network structure information and the graph convolution network cannot be directly applied to the heterogeneous information network,this paper proposes the MG2 vec algorithm.The algorithm first makes an association metric based on the metapath to the network target objects it links to build an association matrix of the target objects.Then,the graph convolution network model is improved,and a convolution operator including an association matrix is introduced.Finally,the convolution operator is convoluted by using propagation rules in the graph convolution network to capture the characteristics of nodes and neighbor nodes,make up for the deficiency of network structure captured by the shallow model,and realize the fusion representation learning of semantic information and structural information.(3)Multi-label classification experiment and visualization experiment are designed and implemented.The experimental results show that the MG2 vec algorithm greatly improves the classification accuracy in the multi-label classification experiment.Compared with the traditional network representation learning algorithm,the classification accuracy of the MG2 vec algorithm on the DBLP data set and the IMDB data set has been increased by 26.99% and 17.77% respectively at most;moreover,the visualization performance is significantly improved,the boundary between different types of nodes is the clearest,and the aggregation of the same type of nodes is also closer.This fully verifies that the low-dimensional representation learned by the MG2 vec algorithm can effectively fuse semantic information and network structure information,and has a better heterogeneous network representation effect.(4)According to the improved algorithm,a paper author classification query system based on B/S architecture is constructed and implemented from the aspects of demand analysis,architecture design,functional module design and database design.Therefore,the graph convolution heterogeneous network representation learning algorithm based on the fusion meta path is applied to the classification system to verify the availability of the algorithm and the practicability of the system. |
PDF Full Text Request