Cross-KG Link Prediction by Learning Substructural Semantics

Link prediction across different knowledge graphs (i.e. Cross-KG link prediction) plays an important role in discovering new triples and fusing multi-source knowledge. Existing cross-KG link prediction methods mainly rely on entity and relation alignment, and are challenged by the problems of KG incompleteness, semantic implicitness and ambiguosness. To deal with these challenges, we propose a learning framework that incorporates both node-level and substructure-level context for cross-KG link prediction. The proposed method mainly consists of a neural-based tensor-completion module and a graph-convolutional-network module, which respectively captures the node-level and substructure-level semantics to enhance the performance of cross-KG link prediction. Extensive experiments are conducted on three benchmark datasets. The results show that our method significantly outperforms the state-of-the-art baselines and some interesting analysis on real cases are also provided in this paper.