Multi-source Information Fusion Based Attributed Heterogeneous Network Embedding

Abstract

Attributed heterogeneous information network embedding is an emerging research topic in the field of network embedding. It can not only embed nodes with complex heterogeneous information into a low-dimensional space, but also preserve the rich attributed information of nodes. Existing works usually only use single-source information of node attributes or network structure, or simply use some shallow models to fuse multi-source information. These methods will lose the original network information and cannot make full use of the rich multi-source information in the attributed heterogeneous network, thus affecting the expression quality of the network embedding. To solve this problem, we propose a multi-source information fusion based attributed heterogeneous network embedding model (MAHE), which improves embedding performances through the deep fusion of heterogeneous structural information and node attribute information. The MAHE model first calculates the attribute similarity of nodes and the node structure similarity calculated by different meta-paths. Then the node attribute similarity matrix is weighted and fused with multiple heterogeneous similarity matrices to obtain multiple multi-source information similarity matrices. This aims to complement the attribute information with the semantic information of different meta-paths. The autoencoder is then used to learn the embeddings of nodes in each multi-source information similarity matrix. Finally, the attention mechanism is used to fuse these node embeddings to learn the importance of different multi-source embeddings.Experiments on three datasets show the superiority and effectiveness of the MAHE model.