KG Completion

Knowledge graph completion studies how to infer missing facts, such as an entity linked by a given relation, from a graph whose recorded knowledge is incomplete. Common approaches include geometric embedding models, graph neural networks and path-based reasoning, rule learning, and methods that add text, images, time, or large language models. Work increasingly targets inductive transfer to unseen entities or graphs, as well as multimodal, temporal, and hyper-relational facts, and is tested on link-prediction benchmarks such as FB15k-237 and WN18RR alongside specialized transfer benchmarks. A central challenge is that an unrecorded fact may be true rather than false, so high ranking accuracy does not by itself guarantee logical faithfulness, reliable confidence, or robust performance on new cases.