Department of Automation, Tsinghua University, Beijing, China.
School of Electronic and Information Engineering, Shenzhen University, Shenzhen, China.
Neural Netw. 2024 Jul;175:106146. doi: 10.1016/j.neunet.2024.106146. Epub 2024 Feb 1.
A Temporal Knowledge Graph (TKG) is a sequence of Knowledge Graphs (KGs) attached with time information, in which each KG contains the facts that co-occur at the same timestamp. Temporal knowledge prediction (TKP) aims to predict future events given observed historical KGs in TKGs, which is essential for many applications to provide intelligent analysis services. However, most existing TKP methods focus on entity and relation prediction tasks but ignore the importance of time prediction tasks. Furthermore, there is uncertainty in time prediction, and it is difficult for prediction models to model it completely. In this work, we propose a collaboration framework with Bayesian Hypernetwork and Time-Difference Evolutional Network (BH-TDEN) to address these problems. First, we begin with the time prediction task, and we present a Bayesian hypernetwork to model the uncertainty of events time. For the input of Bayesian hypernetwork, we design a novel time-difference evolutional network to obtain the entities and relations embedding. Specifically, we propose an auto-regressive time gate parameterized by the time difference of adjacent KGs in entity and relation encoder to learn the time-sensitive TKG embedding, which not only learns the relationship between the given time information and TKG embedding but also provides more expressive TKG embedding for Bayesian hypernetwork to accurately predict the time of future events. Furthermore, we also present a novel relation updating mechanism that employs the neighbor relations of the subject corresponding to the current relation to learn more adaptive relation embedding. Extensive experiments demonstrate that the proposed method obtains considerable time prediction and link prediction performance on four TKG benchmark datasets.
时态知识图 (TKG) 是附有时间信息的知识图 (KG) 的序列，其中每个 KG 包含在同一时间戳同时发生的事实。时态知识预测 (TKP) 旨在根据 TKG 中观察到的历史 KGs 预测未来事件，这对于许多应用程序提供智能分析服务至关重要。然而，大多数现有的 TKP 方法侧重于实体和关系预测任务，但忽略了时间预测任务的重要性。此外，时间预测存在不确定性，预测模型很难完全对其进行建模。在这项工作中，我们提出了一个具有贝叶斯超网络和时间差分进化网络 (BH-TDEN) 的协作框架来解决这些问题。首先，我们从时间预测任务开始，提出了一个贝叶斯超网络来建模事件时间的不确定性。对于贝叶斯超网络的输入，我们设计了一种新颖的时间差分进化网络来获取实体和关系嵌入。具体来说，我们提出了一种自动回归时间门，由相邻 KG 之间的时间差参数化，用于在实体和关系编码器中学习时间敏感的 TKG 嵌入，这不仅学习了给定时间信息与 TKG 嵌入之间的关系，还为贝叶斯超网络提供了更具表现力的 TKG 嵌入，以准确预测未来事件的时间。此外，我们还提出了一种新颖的关系更新机制，利用与当前关系对应的主体的邻居关系来学习更具适应性的关系嵌入。广泛的实验表明，所提出的方法在四个 TKG 基准数据集上获得了相当的时间预测和链接预测性能。
