Abstract: Knowledge reasoning is a fundamental task in knowledge graph completion. It has received widespread attention from the academic community. With the development of knowledge reasoning technology, applying graph neural networks to knowledge reasoning methods can fully consider the structural information of knowledge graphs, making them more interpretable and stronger in reasoning ability, which is currently one of the research hotspots. The basic concepts of knowledge graph and knowledge reasoning are described. Knowledge reasoning methods based on graph neural networks are categorized from the perspectives of closed-world and open-world settings. In the closed-world context, the emphasis is placed on two types of methods: graph convolutional networks and graph attention networks. In the open-world context, semi-inductive and fully-inductive methods are explored. The typical model frameworks of these methods are compared and analyzed, and their respective strengths and weaknesses are summarized. Finally, the applications of graph neural network reasoning in intelligent question answering, recommendation system and biomedicine are discussed, and the future research direction of graph neural network based knowledge reasoning is prospected.

Key words: knowledge graph, graph neural networks, knowledge reasoning, closed world, open world, inductive reasoning

摘要： 知识推理作为知识图谱补全中一项重要任务，受到了学术界的广泛关注。随着知识推理技术的发展，将图神经网络应用于知识推理的方法可以充分考虑知识图谱的结构信息，使其具备更好的可解释性和更强的推理能力，是目前的研究热点之一。对知识图谱和知识推理的基本概念进行了叙述。从封闭世界和开放世界两个角度对基于图神经网络的知识推理方法进行了归纳。在封闭世界下，介绍了图卷积神经网络和图注意力神经网络两类方法；在开放世界中，探讨了半归纳式和全归纳式两种方法。对这些方法的典型模型框架进行了对比分析，并总结了各自的优缺点。最后对图神经网络推理在智能问答、推荐系统以及生物医疗上的应用进行了探讨，并对基于图神经网络知识推理的未来研究方向进行了展望。

关键词: 知识图谱, 图神经网络, 知识推理, 封闭世界, 开放世界, 归纳推理