Heterogeneous Network Representation Learning Method Based on Meta-path

Abstract

Mining the rich structure and semantic information hidden in heterogeneous information networks is one of the important tasks of network representation learning. At present, there are relatively few studies on network representation for heterogeneous information network which contains different types of nodes and link relationships. Most studies on network science are based on the homogeneous networks where nodes are objects of the same entity type and the links between nodes are also of the same type. In this paper, we propose a new network representation learning method for heterogeneous network. The core of this method contains tow parts: First, according to semantic of different meta-paths, we get weights between nodes of the same type in heterogeneous information network based on attention mechanism. And then, through biased random walk on nodes of the same type, we get node sequences which can be processed by the skip-gram model to generate representation vectors of nodes. To verify our method, we do experiments on DBLP and Aminer dataset. The experimental results demonstrate that the embeddings learned from the proposed model achieve better performance than state-of-the-art methods in tasks including node classification and similarity research.