FTPComplEx: A flexible time perspective approach to temporal knowledge graph completion

Abstract

The dynamic nature of interconnected data evolving over time poses significant challenges for graph representation and reasoning, particularly as temporal knowledge graphs scale in size and complexity. Existing models like TPComplEx (Time Perspective Complex Embedding) leverage tensor decomposition techniques to capture temporal dynamics, but their static weighting approach often lacks the flexibility needed to adapt to the nuanced evolution of relationships and entities. This rigidity can lead to missed temporal dependencies and loss of valuable insights, especially in large-scale graphs comprising millions or even billions of factual entries. To overcome these limitations, we propose FTPComplEx (Flexible Time Perspective Complex Embedding), a novel embedding model that introduces adjustable weights to dynamically modulate the influence of temporal information. This flexibility enables FTPComplEx to more accurately capture the intricate interactions between entities, relations, and time, providing a more robust understanding of temporal dynamics within knowledge graphs. Our extensive evaluations on benchmark datasets, including YAGO15k, ICEWS, and GDELT, demonstrate that FTPComplEx achieves state-of-the-art results, outperforming TPComplEx and other existing models. Notably, on the YAGO15k dataset, FTPComplEx achieves a 9.04% improvement in Mean Reciprocal Rank (MRR) and an 11.35% increase in Hits@1, demonstrating its effectiveness in managing complex temporal relationships. Further analysis shows that FTPComplEx maintains strong performance even with lower-rank embeddings, significantly reducing computational costs while maintaining accuracy.