In recent years, the knowledge graph representation learning has been used to represent the components of the knowledge graphs in a low-dimensional vector embedding, as a mainstream way to combine the artificial intelligence with the knowledge graphs. This paper reviews the mainstream knowledge graph representation learning methods without auxiliary information, mainly, the distance-based and the semantic matching-based methods, and the knowledge graph representation learning methods containing textual auxiliary information and category auxiliary information, along with the advantages and the disadvantages of various representation learning methods. It is found that the introduction of auxiliary information can effectively represent new entities and relationships in the knowledge graph, but the time and space costs are significantly increased, and thus the methods without auxiliary information are more easily applied in practical scenarios at this stage. Finally, we show how the knowledge graph embedding can be applied to downstream tasks such as the triad classification, the link prediction and the recommender systems. A collection of datasets and open source libraries for different tasks is compiled and, and a comprehensive outlook on promising research directions such as large-scale, dynamic knowledge graphs is given.

knowledge graph  /  knowledge graph embedding  /  representation learning  /  deep learning