Time Split Network for Temporal Knowledge Graph Completion

Temporal Knowledge Graphs (TKGs), represented by quadruples, describe facts with temporal relevance. Temporal Knowledge Graph Completion (TKGC) aims to address the incompleteness issue of TKGs and has received extensive attention in recent years. Previous approaches treated timestamps as a single node, resulting in incomplete parsing of temporal information and the inability to perceive temporal hierarchies and periodicity. To tackle this problem, we propose a novel model called Time Split Network (TSN). Specifically, we employed a unique approach to handle temporal information by splitting timestamps. This allows the model to perceive temporal hierarchies and periodicity, while reducing the number of model parameters. Additionally, we combined convolutional neural networks with stepwise fusion of temporal features to simulate the hierarchical order of time and obtain comprehensive temporal information. The experimental results of entity link prediction on the four benchmark datasets demonstrate the superiority of the TSN model. Specifically, compared to the state-of-the-art baseline, TSN improves the MRR by approximately 2.6% and 1.3% on the ICEWS14 and ICEWS05-15 datasets, and improves the MRR by approximately 33.5% and 34.6% on YAGO11k and Wikidata12k, respectively.