Hierarchical Logic-Driven Relational Learning for Knowledge Graph Completion

Knowledge graphs (KGs), represented as structured head entity-relation-tail entity triples (fact triples), provide a systematic framework for knowledge representation and reasoning across various domains. However, as the scale of KGs expands, challenges such as data sparsity, missing attributes, and incomplete relationships increasingly undermine their effectiveness in tasks like semantic inference and link prediction. Addressing these challenges requires methods capable of capturing and leveraging the latent relational semantics embedded in hierarchical structures. This paper introduces a hierarchical relational learning framework, named HLogic, which emphasizes the structured discovery of hierarchical logic patterns to improve knowledge graph completion. The proposed method integrates embedding-based models with a carefully designed graph neural network that serves as a relational learner, explicitly designed to capture multi-level dependencies through hierarchical logic. By modeling the regularities and patterns inherent in the layered relationships of KGs, HLogic effectively identifies and utilizes the structural and semantic coherence of knowledge graphs. This hierarchical logic-driven approach enhances the consistency and expressiveness of knowledge embeddings. The iterative optimization process further refines the embeddings, preserving both local and global relational semantics while improving link prediction accuracy. Experiments conducted on benchmark datasets demonstrate that HLogic achieves notable performance improvements in metrics compared to baseline methods. These results underscore the importance of explicitly capturing hierarchical logic regularities and highlight the potential of HLogic to adapt to diverse relational structures. This study not only showcases the value of hierarchical logic in relational learning but also lays the groundwork for future explorations into integrating logic-driven methodologies with advanced graph learning techniques for real-world applications.