KGPFN: Knowledge Graph Foundation Model with In-Context Learning

Researchers have introduced KGPFN, a foundational model for knowledge graphs (KG) that combines in-context learning with transferable relational patterns. Unlike traditional approaches that emphasize relation-level universality, KGPFN utilizes both the immediate context surrounding query entities and a broader context that summarizes relational behavior across various instances. It employs a Prior-data Fitted Network to integrate these elements, initially learning relation representations through message passing on relation graphs to capture invariances across graphs, followed by encoding local neighborhoods for reasoning specific to queries. This method tackles the often-overlooked significance of in-context learning in KG reasoning, with the goal of generalizing across graphs containing previously unseen entities and relations.