Leveraging Static Models for Temporal Knowledge Graph Completion

Abstract

The inclusion of temporal information in knowledge graph embedding (KGE) has remained relatively unexplored, even though it stands to argue that this would result in better embeddings. Additionally, models that do include the temporal component perform only marginally better than those that do not (static models). Noting this, we introduce SpliMe, a model-agnostic pre-processor for temporal knowledge graphs that makes it possible to embed them with static models. We show that SpliMe achieves state-of-the-art performance on two datasets commonly used for temporal KGE method evaluation and increases performance with regards to our baseline on another dataset. Furthermore, we uncover problems with existing evaluation procedures for static KGE models on temporal graphs and propose a simple method to fix these issues. Finally, we redefine the link prediction metric for temporal knowledge graph embedding to better suit temporal scope prediction.