Learning Rules in Knowledge Graphs via Contrastive Learning

In this paper, we study the problem of learning logical rules for reasoning in KGs. Logical rules have the ability to enhance KG reasoning and offer interpretability. Recent rule learning approaches for large-scale KGs encounter two distinct challenges: high computational complexity resulting from intensive matrix multiplications, or a bias towards syntax (structure) over semantic features, which is often manifested by a narrow focus on the “close-path” of rules. To address these challenges, we propose CLRL, a model that utilizes contrastive learning to capture deeper semantic features of paths in KGs to learn high-quality rules. To further enhance performance and make contrastive learning more adapted to the rule learning task, we propose a deductive dropout strategy for data augmentation. Finally, we evaluate the effectiveness and efficiency of CLRL on large-scale KGs and demonstrate that CLRL achieves state-of-the-art results in KG completion tasks.