Heterogeneous Information Network Embedding with Meta-path Based Graph Attention Networks

Abstract

Network embedding is an emerging research field which aims at projecting network elements into lower dimensional spaces. However, most network embedding algorithms focus on homogeneous networks, thus cannot be directly applied to the Heterogeneous Information Networks (HINs) which are prevalent in real world systems. Therefore, how to effectively preserve both the structural and semantic information, as well as higher-level proximity of large-scale HINs still remains an open problem. In this paper, we propose a novel heterogeneous network embedding model called Meta-path based Graph ATtention nEtwork (MetaGATE). To tackle the problem of preserving both structural and semantic information for HINs, we adopt a multiple meta-paths based random walk scheme with Skip-Gram for sampling and pre-training. In addition, the model involves graph attention networks to collect and aggregate heterogeneous information, which may reveal higher-level implicit semantics. Besides, to ensure stability, multi-head attention mechanism is employed in the graph attention model by concatenating the embedding of multiple independent self-attention processes. Experiments on two real-world bibliography networks are conducted. Compared to state-of-the-art baselines, our method achieves the best performance on the node classification task, which shows the effectiveness of the proposed model.