PEARL Enhances Inductive Knowledge Graph Completion with Context.

Inductive knowledge graph completion (IKGC) is a challenging task that involves predicting missing links for entities not encountered during model training, requiring the learning of transferable relational and structural patterns. Existing methods often treat relational paths in isolation from their surrounding structural context, even though a path's predictive relevance can vary significantly based on its environment. The PEARL (Path-Entity Aligned Relational Learning) framework addresses this by modeling paths as reasoning signals that are conditioned by their context.PEARL constructs a unique "contextual subgraph" for each query, formed by combining the neighborhoods of the query entities. It then employs a large language model (LLM)-guided retriever to identify and distill the most semantically relevant paths within this subgraph. A bipartite interaction graph is subsequently built, connecting these paths with contextual entities and a global subgraph representation. This architecture allows path embeddings to dynamically adapt to both local and global structural evidence, providing a more nuanced understanding of relationships.To counter potential noise from the expanded context, PEARL incorporates a dual-view contrastive objective, which promotes consistent representations even under stochastic perturbations of the context. Experimental results on standard benchmarks (WN18RR, FB15k-237, NELL-995) demonstrate that PEARL achieves superior average Hits@10 compared to other IKGC methods. Ablation studies and efficiency analyses further validate the contributions of its contextual subgraph modeling, semantic path retrieval, path-entity interaction, and contrastive regularization components.