Subgraph-aware Few-Shot Inductive Link Prediction via Meta-Learning

Link prediction for knowledge graphs aims to predict missing connections between entities. Prevailing methods are limited a transductive setting and hard process unseen The recently proposed subgraph-based models provide alternatives links from the subgraph structure surrounding candidate triplet. However, these require abundant known facts of training triplets perform poorly on relationships that only have few triplets. In this paper, we propose Meta-iKG, novel meta-learner few-shot inductive relation reasoning. Meta-iKG utilizes local subgraphs transfer subgraph-specific information rapidly learn transferable patterns via meta-gradients. way, find model can quickly adapt using handful with settings. Moreover, introduce large-shot updating procedure ensure our generalize well both relations. We evaluate benchmarks sampled NELL Freebase, results show outperforms currently state-of-the-art in scenarios standard