Temporal knowledge graph reasoning triggered by memories

The task of inferring missing facts using the temporal knowledge graph (TKG) is important and has been widely studied. Extrapolation in TKG inference is more challenging because no direct historical facts are available for predicting future events. Previous methods apply recursive embedding learning to solve the extrapolation problem. However, these methods primarily focus on recent historical facts and overlook the potential knowledge hidden in earlier facts. Moreover, the recursive embedding learning process can lead to imprecise knowledge propagation. To address these issues, we propose a memory-triggered decision-making (MTDM) network. It takes advantage of earlier historical facts to establish initial node representations and then updates node representations with recent historical facts. In particular, to enhance the prediction performance and efficiency, we update node representations of each timestamp in parallel using the proposed Res-GCN and then establish temporal correlations among these representations based on the designed control gating unit. To mitigate the issue of imprecise knowledge propagation, we leverage the facts most relevant to the missing knowledge to directly make predictions. This strategy improves the predictions by focusing on the most informative and contextual-free information. Additionally, we introduce the dissolution constraint, which enables us to analyse the process of event dissolution. Extensive experiments demonstrate that MTDM outperforms existing methods in terms of prediction accuracy and computational efficiency.