Improving knowledge graphs via data-based causal structures

Derui Lyu, Xiangyu Wang, Lyuzhou Chen, Taiyu Ban, Kaiming Xiao, Lihua Liu

Abstract: Knowledge graphs, prevalent in a multitude of domains, have gained significant traction over the years. However, these graphs, often created from noisy sources using imperfect methods, suffer from inaccuracies, inconsistencies, and incompleteness. The imperative to rectify and supplement these issues is frequently addressed by costly domain-expert-based strategies or learning-based techniques, which can prove unreliable in noisy environments. Real-world data, with its inherent objectivity and implicit knowledge, can serve as an external source of knowledge. The ready availability of big data today presents an opportunity for improving knowledge graphs. Yet, the challenge of extracting latent information from such data has often deterred the inclusion of numerical data in current research. This paper introduces a pioneering approach that leverages external numerical data to enhance knowledge graph quality. The proposed method commences by mining statistical relationships, such as correlation and causality from the data. Following a thorough analysis of causal patterns and an evaluation of relation strengths, it identifies and eliminates redundant knowledge. The empirical results from numerous experiments validate the efficacy of our approach. The proposed method outperforms existing learning-based methods, demonstrating superior accuracy and stability in knowledge correction.

External IDs:dblp:journals/kais/LyuWCBXL25