Few-Shot Relational Learning on Knowledge Graphs: Towards Model Adaptation and Generalization

Abstract

Knowledge graphs are essential for structuring vast information and enabling advanced inference in applications like QA, web search, and recommendation. However, their inherent incompleteness limits reasoning, driving research in relational learning to infer missing facts via expressive representations. Traditional embedding-based methods succeed but rely on large-scale data. Many relations, however, have few triplets, making generalization difficult. Few-shot relational learning addresses this by learning from minimal examples. Despite progress, challenges remain: (1) overemphasis on entity embeddings over relational structures, (2) assumptions violating permutation invariance, (3) task-isolated learning that misses transferable patterns, and (4) reliance on relational data while ignoring semantic knowledge. This thesis systematically tackles these challenges. First, we introduce a hierarchical framework that jointly models entity, triplet, and context information, ensuring permutation invariance and improved generalization. Second, we propose a meta-learning framework with a Mixture-of-Experts model for relational prototypes, balancing global generalization and local adaptability. Third, we integrate semantic knowledge via a prompted meta-learning framework, enhancing inference of unseen relations and paving the way for large language models. Additionally, we release pretrained semantic embeddings for benchmark datasets to support future research. This thesis lays a foundation for more robust and semantically enriched knowledge graph relational learning, advancing intelligent knowledge graphs.