Neighboring relation enhanced inductive knowledge graph link prediction via meta-learning

Inductive link prediction over knowledge graphs(KGs) aims to perform inference over a new graph with unseen entities. In contrast to transductive link prediction, which learns a fixed set of known entities and cannot handle unseen entities, inductive link prediction is closer to real-world scenarios because of the evolving nature of KGs. The recently proposed methods utilize the enclosing subgraph of candidate triples to obtain inductive ability. However, these methods explicitly encode the local subgraph surrounding each node pair and neglect the entity-independent neighboring relation modeling, resulting in poor generalization to unseen entities and high time complexity. Inspired by the success of meta-learning on out-of-domain generalization, we propose a novel N eighb O ring relati O n enhance D meta- L earn E r(NOODLE) for inductive link prediction. NOODLE sufficiently exploits the neighboring relations from two aspects: commonsense semantic features and topological correlation information. Moreover, NOODLE rapidly learns transferable relations-specific information via meta-learning and generalizes well to unseen entities. Extensive experiments show that NOODLE outperforms the state-of-the-art methods on commonly used benchmark datasets in terms of effectiveness, generalization, and efficiency.