| 摘要: |
| 知识超图是一种使用多元关系表示现实世界的异构图,但无论在通用领域还是垂直领域,现有的知识超图普遍存在不完整的情况.因此,如何通过知识超图中已有的链接推理缺失的链接,是一个具有挑战性的问题.目前,大多数研究使用基于多元关系的知识表示学习方法完成知识超图的链接预测任务,但这些方法仅从时间未知的超边中学习实体与关系的嵌入向量,没有考虑时间因素对事实动态演变的影响,导致在动态环境中的预测性能较差.首先,根据首次所提出的时序知识超图定义,提出时序知识超图链接预测模型,同时从实体角色、位置和时序超边的时间戳中学习实体的静态表征和动态表征,以一定比例融合后作为实体嵌入向量用于链接预测任务,实现对超边时序信息的充分利用.同时,从理论上证明模型具有完全表达性和线性空间复杂度.此外,通过上市公司的公开经营数据构建时序知识超图数据集CB67,并在该数据集上进行了大量实验评估.实验结果表明,模型能够在时序知识超图数据集上有效地执行链接预测任务. |
| 关键词: 时序知识超图 链接预测 知识表示 嵌入学习 时序信息 |
| DOI：10.13328/j.cnki.jos.006888 |
| 分类号: |
| 基金项目:科技创新2030“新一代人工智能”重大专项(2020AAA0108504);国家自然科学基金(61972275) |
|
| Towards Time-aware Knowledge Hypergraph Link Prediction |
|
CHEN Zi-Rui1,2, WANG Xin1,2, WANG Chen-Xu1,2, ZHANG Shao-Wei1,2, YAN Hao-Yu1,2
|
|
1.College of Intelligence and Computing, Tianjin University, Tianjin 300350, China;2.Tianjin Key Laboratory of Cognitive Computing and Application, Tianjin 300350, China
|
| Abstract: |
| A knowledge hypergraph is a form of a heterogeneous graph that represents the real world through n-ary relations. However, both in general and specific domains, existing knowledge hypergraphs often suffer from incompleteness. Therefore, it is a challenging task to reason the missing links through the existing links in the knowledge hypergraph. Currently, most research employs knowledge representation learning methods based on n-ary relations to carry out link prediction tasks in knowledge hypergraphs. However, these methods only learn embedding vectors of entities and relations from hyperedges with unknown temporal information, neglecting the impact of temporal factors on the dynamic evolution of facts, resulting in poor predictive performance in dynamic environments. Firstly, based on the definition of temporal knowledge hypergraph that proposed by this study for the first time, a link prediction model is proposed for temporal knowledge hypergraphs. Simultaneously, static and dynamic representations of entities are learnt from their roles, positions, and timestamps of temporal hyperedges, which are merged in a certain proportion and utilized as final entity embedding vectors for link prediction tasks to realize the full exploitation of hyperedge temporal information. At the same time, it is theoretically proved that the proposed model is fully expressive and has linear space complexity. In addition, a temporal knowledge hypergraph dataset CB67 is constructed from the public business data of listed companies, and a large number of experimental evaluations are conducted on this dataset. The experimental results show that the proposed model can effectively perform the link prediction task on the temporal knowledge hypergraph dataset. |
| Key words: temporal knowledge hypergraph link prediction knowledge representation embedding learning temporal information |
