Attention Based Meta Path Fusion for Heterogeneous Information Network Embedding

Abstract

Recently, there is a surge of network embedding algorithms, which embed information network into a low dimensional space. However, contemporary network embedding algorithms focus on homogeneous networks, while we know that many real-world systems can be constructed with heterogeneous information networks (HINs). Compare to homogeneous networks, HINs contain heterogeneity types of nodes and edges, which leads to new challenges for traditional network embedding: handing mixed heterogeneous nodes and fusing rich semantic information. Although several HIN embedding algorithms have been proposed, these challenges have not been well dressed. How to explore the rich semantic information and integrate these information still remain to be solved. In this paper, we propose a novel attention based meta path fusion model for HIN embedding (called AMPE). In order to handle node heterogeneity and extract rich information, AMPE first extracts multiple homogeneous networks from HIN with meta paths, and then employs adopted AutoEncoders to embed these homogeneous networks. After that, AMPE fuses these embeddings learned from homogeneous networks with attention mechanism. Experimental results on two real-world datasets demonstrate the effectiveness of the proposed model.