Heterogeneous Information Network Representation Learning Algorithm Based On Transition Probability Matrix Of Meta-path

Objects and the connection of objects constitutes a network,people can see such networks everywhere in daily life,such as social networks,road networks,biological networks,the World Wide Web,language networks,etc.The analysis of these networks has very practical research and application value,such as node classification,recommendation system,link prediction,network recovery,biomedical,etc.Due to the scale problem,it is not convenient to directly calculate and store these networks,thereby network representation learning emerges.Network representation learning is also known as network embedding or graph embedding.Network representation learning projects the original high-dimensional,sparse network into a low-dimensional,dense space and retains some relevant network characteristics according to relevant application tasks.A good network representation learning algorithm can preserve the local and global characteristics of nodes in the original network,traditional network representation learning algorithms generally use the adjacency matrix to calculate directly or indirectly,and the matrix calculation has high time and space complexity.So the problem of traditional network representation learning algorithms are sparse,high-dimensional,difficult to calculate and unable to express semantics.How to effectively represent the network has two main challenges: large-scale(millions of nodes and billions of edges in the network);heterogeneity(heterogeneous information networks closer to the real world have multiple types of nodes and different types of edges).The main points and innovations of this paper are as follows:1.This paper proposes a new propagation probability matrix calculation method for heterogeneous information networks.It takes advantage of the adjacency matrix and the meta path.Traditional network representation learning mostly targets homogeneous information networks.Nodes and the relationship of nodes in homogeneous information networks is the same type.Nodes and the relationship of nodes in heterogeneous information networks is different types,nodes and the relationship of nodes have different meanings.It is obviously unreasonable to directly use the node similarity index in the homogeneous network without distinguishing these meanings.This paper proposes a new transition probability matrix calculation method for heterogeneous information networks,which utilizes adjacency matrix and meta-path.2.This paper proposes a new heterogeneous information network node similarity index.To represent the network involves how to design node similarity index.In this paper,a new heterogeneous information network node similarity index is proposed,which combines the first-order second-order similarity of nodes and the transition probability matrix.It is proved by experiments that the similarity index of this paper is better than the comparison node similarity index in the data set used in the experiment.3.Propose an algorithm HINtpm for the characteristics of heterogeneous information networks.Aiming at the characteristics of heterogeneous information networks,an algorithm HINtpm is proposed,which uses the previously proposed heterogeneous information network node similarity index to measure the similarity between heterogeneous information network nodes,then use the automatic encoder to reduce the dimension to get the final node representation,it is displayed on the three application tasks of classification,clustering and visualization.The experimental results show that the heterogeneous information network representation learning algorithm proposed in this paper has different degrees of improvement in application tasks.