Global-Local Latent on Learning Model for Temporal Knowledge Graphs Based on Graph Neural Networks

Temporal Knowledge Graphs (TKGs) are an effective method for representing inter-entity and inter-event relationships in continuous time. The inference of temporal knowledge graphs refers to the completion of missing information for already existing points in time, and the extrapolation is the prediction of unknow-able facts occurring in the future, and these two kinds of reasoning have important practical value in various fields. However, in the reasoning of temporal knowledge graphs, the lack of capturing potential relationships between entities between dif-ferent times or within the same time when encoding local and global historical information has led to models that often perform poorly in the face of complex political events. To address these issues, this paper proposes a global-local latent relationship learning model for temporal knowledge graphs based on graph neural networks. Specifically, the model employs a latent relationship learning module to mine and exploit latent relationships within a time slice with multiple adjacent time slices to better simulate and interpret local and global historical information, enhancing the model's ability to deal with complex real-world event problems. Experimental results on four benchmark datasets show a 3 to 4 percentage point improvement in the model's inference capability compared to the state-of-the-art inference model.