Domain Information Learning For Multi-Domain Knowledge Graph Link Prediction

The concept of instance changes with domains in a multi-domain knowledge graph. Methods for Link prediction typically focus on embedding the instances in the knowledge graph to infer missing links. However, multi-channel graph exists as multi-concept entities in which embedding entities and relations on an instance level may not be able to satisfy complete concepts. In addition, the methods with the usage of outer information are biased by the domain information. To address the issues, we propose a new bi-directional domain learning method to learn the representation of instances on a concept level with latent domain information in the multi-domain knowledge graphs. It consists of bi-directional sampling, multi-domain extension, and multi-channel attention to learn and extract concepts from domain knowledge graphs. The experimental results of link prediction show that our approach performs well by meeting the required criteria.