Temporal knowledge graph forecasting query based on global-local historical information

Abstract

Temporal knowledge graph (TKG) queries aim to retrieve relevant facts that conform to time constraints to answer a given query by reasoning known TKG facts. The continuous development of TKG query research has extended TKG queries to the TKG forecasting domain, enabling the forecasting of answers to unknown queries by leveraging historical information from query questions. However, TKG forecasting query research is currently facing two considerable challenges. Firstly, existing TKG forecasting query methods cannot adequately capture the global historical information of query questions, which makes it difficult to effectively mine periodic features, repetitive patterns, and dynamic evolution characteristics of new events. Secondly, when modeling local historical information, existing methods fail to focus on the historical correlation of facts between adjacent timestamps, ignoring the crucial role of local information in the temporal evolution process. In this paper, a TKG forecasting query framework based on global-local historical information is proposed to solve the above challenges. Specifically, for the global historical information of the query question, the periodic and repetitive patterns of historical facts and the potential changing laws of non-historical facts are learned by modeling global historical facts and non-historical facts. Concerning the local historical information, entities and relations are aggregated in knowledge graph (KG) snapshots and their changes and evolution are simulated at adjacent timestamps to enhance the ability of the model to capture temporal dependencies. At the same time, the impact of local snapshots on query questions is quantified to capture the evolution process of local information more accurately. Finally, we design dedicated scoring functions for different types of query tasks to achieve effective query forecasting. Extensive experiments on four datasets demonstrate that the proposed model has better performances in forecasting unknown queries than other baseline models.