HEAM: Heterogeneous Network Embedding with Automatic Meta-path Construction.

Heterogeneous information network (HIN) embedding is widely used in many real-world applications. Meta-path used in HINs can effectively extract semantic information among objects. However, the meta-path faces challenges on the construction and selection. Most of the current works construct dataset-specific meta-paths manually, which rely on the prior knowledge from domain experts. In addition, existing approaches select a few explicit meta-paths, which lack of much subtle semantic information among objects. To tackle the problems, we propose a model with automatic meta-path construction. We develop a hierarchical aggregation to learn effective heterogeneous embeddings with meta-path based proximity. We employ a multi-layer network framework to mine long meta-paths based information implicitly. To demonstrate the effectiveness of our model, we apply it to two real-world datasets and show the performance improvements over state-of-the-art methods.