<p>Inductive link prediction (ILP) in knowledge graphs (KG) is gaining significant attention, focusing on predicting missing triples involving unseen entities during training. While subgraph-based methods are prevalent, they often encounter two primary limitations: (i) a tendency to concentrate on fully-inductive setting with unseen-unseen entities, neglecting the challenge of a more realistic truly-inductive setting involving both unseen entities and unseen relations, and (ii) a constraint in handling local subgraphs, overlooking the global structural information of the KG. To address these challenges, we propose RPMI, a novel model utilizing a relational message passing network with mutual information maximization tailored for the truly-inductive setting. Specifically, RPMI extracts enclosing and one-hop disclosing subgraphs around target triples, incorporates topological patterns between relations to transform entity graphs into relational graphs for subgraph reasoning, and introduces techniques like injecting KG’s ontological schema and relation-aware neighborhood attention. Moreover, to enhance the global modeling of topological patterns between relations, we maximize subgraph-graph interaction information. Extensive experiments on various inductive benchmark datasets demonstrate RPMI’s substantial performance improvement over existing methods in both fully and truly-inductive link predictions.</p>