Inductive link prediction via global relational semantic learning

Knowledge graphs (KGs) play a crucial role in storing and utilizing real-world facts, but they often suffer from sparse and missing relations. To overcome these challenges, researchers have proposed relation prediction models, including embedding-based methods. However, these methods are restricted to the transductive setting and require retraining when new entities emerge. Thus, recent research has focused on the inductive setting, allowing for different entities in the test set. Subgraph-based models utilizing graph neural networks (GNNs) for local structural information aggregation have shown promising performance. However, existing approaches focus only on local structural information, ignoring the semantic correlation among relations in the global perspective, resulting in sub-optimal performance. Thus, we propose an inductive relation prediction model GRelGT that incorporates the global relation graph with topological information and the enclosing subgraph. GRelGT consists of two core components: a global relation graph module and a subgraph module. The global relation graph module converts the original knowledge graph into a relation graph, with nodes representing edges (triples) in KGs. Furthermore, we introduce four topological structural features as edge types in the global relation graph to facilitating the learning of the semantic correlations between relations. By leveraging the topological features of the relations, the model’s ability to capture the hidden patterns in the KG is enhanced. Meanwhile, the subgraph module is dedicated to exploring the local structural and semantic information within the enclosing subgraph around the target triple. For a more precise understanding of semantic correlations, we further introduce global relation-aware attention and local query-aware attention mechanisms in the subgraph GNN. This allows GRelGT to dynamically weigh the importance of different relations, effectively leveraging both global and local information for inference. Experimental results on three KG datasets demonstrate the superiority of our model compared to state-of-the-art approaches.