Multi-dimension rotations based on quaternion system for modeling various patterns in temporal knowledge graphs

Abstract

Missing information is a prevalent occurrence in temporal knowledge graphs (TKGs), and thus, TKG completion holds considerable importance. Modeling the diverse relational patterns inherent in TKGs is crucial for this process. However, existing methods mainly focus on pre-existing patterns within knowledge graphs while neglecting the influence of temporal information. It is common for multiple relationships to exist between two entities at the same moment, as well as for the same event to transpire at different timestamps. Existing models primarily rely on single or sequential transformations, rendering them inadequate for modeling these intricate patterns. To tackle these challenges, we propose a novel model, multi-dimension rotations based on quaternion system (MDRQS), that integrates the attention mechanism to fuse rotations of different dimensions for modeling interactions between entities. This complex combination of transformations, utilizing parallelization, enables the modeling of the aforementioned patterns through multi-dimensional rotations. The attention mechanism determines the most appropriate dimensional rotation for different facts at various timestamps. We demonstrate that MDRQS effectively models pre-existing and new patterns. Through experiments conducted on four benchmark datasets, the effectiveness of our model is shown in the link prediction task.