Potential subgraph rule and reasoning context enhancement for sparse multi-hop knowledge graph reasoning

Abstract

Multi-hop knowledge graph reasoning aims to leverage the relations between multiple nodes in a knowledge graph to reason information about an event or entity. This reasoning process requires traversing multiple interconnected facts or knowledge points, which aids in understanding the model’s decision-making process. Multi-hop knowledge graph reasoning has driven the development of knowledge-based technologies, such as question-answering systems and recommendation systems. However, multi-hop reasoning relies on the connectivity between different entities in the knowledge graph. This characteristic makes multi-hop reasoning lack robustness when dealing with sparse data. To address the challenges of sparsity, recent studies pre-train knowledge graph embedding models to complete potential triples. The completion methods introduce noisy triples, which increases the risk of model selection errors and spurious paths. In this work, we propose a framework based on potential subgraph rule and reasoning context enhancement to mitigate the challenges of sparsity. On one hand, we leverage reasoning context to enhance state information and the reasoning process; on the other hand, we design an action perceptron based on the importance of reasoning context to reduce the introduction of noisy triples. Additionally, we analyze the phenomenon of data augmentation introducing spurious paths, and further utilize data augmentation-based potential subgraph rules to guide the reasoning process. This dual mechanism demonstrates stronger robustness in addressing sparsity challenges and spurious paths. Diverse experiments demonstrate that our model outperforms the existing multi-hop reasoning models across five datasets. Our implementations will be publicly available at: https://github.com/jianruichen/PreKGR.