Method And Application Research Of Heterogeneous Network Representation Learning Via Neural Network

The representation learning method of meta path random walk on heterogeneous network can effectively obtain the semantic information between different types of nodes and improve the quality of node representation.There are obvious shortcomings on discrimination due to the neglect of the global structure information in heterogeneous networks.This paper focuses on the fusion of local structure information and global clustering information,the main contents are as follows:Metapath2vec and Metapath2vec++,the representation learning methods of heterogeneous networks,maintain the local nearest neighbor structure between nodes,and do not consider the global clustering structure of heterogeneous networks,so as to reduce the accuracy of node representation in networks.To solve this problem,this article proposes two kinds of heterogeneous network representation learning models: HINSC and HINSC++.Firstly,the meta path random walk strategy is used to obtain the node sequence including the local nearest neighbor structure and semantic information,and the one-hot representation corresponding to the node is used as the input of the feedforward neural network.After the nonlinear transformation of hidden layer,k-means clustering is carried out for nodes of the same type in heterogeneous networks by adding clustering constraint information to the objective function,so that the nodes in the output layer can maintain the local nearest neighbor structure and global clustering structure at the same time.The low dimensional vector representation of nodes in heterogeneous networks is learned by using stochastic gradient descent algorithm.The learned vector representation is used for classification and clustering experiments on two data sets of heterogeneous academic networks,aminer and DBLP.Experimental results show that the NMI of the vector representation trained by HINSC and HINSC++ models is increased by 12.46%?26.22% in clustering task,and 9.32%?17.24% in classification task,which further verifies the validity of the model.In order to explore the practical application value of the proposed algorithm,on the premise of not change the original semantic information of heterogeneous academic network,this experimental extracted two different types of nodes and their link relations,and reconstructed the heterogeneous information network on this basis.According to the proposed HINSC and HINSC++ models in this paper,the vector representation of two different types of nodes in the network is trained,by calculating the cosine similarity between node vector representation to represent the similarity between nodes,so as to predict the possibility of future links between two types of nodes in the network.The experimental results show that the AUC of HINSC and HINSC++ models in link prediction task increases by 9.3%?20.3%,which verifies the effectiveness of the proposed model in heterogeneous academic network link prediction task.