TS-align: A temporal similarity-aware entity alignment model for temporal knowledge graphs

Entity Alignment (EA) is a crucial step in knowledge graph fusion, aiming to match equivalent entity pairs across different knowledge graphs (KGs). In recent years, Temporal Knowledge Graphs (TKGs) have extended static KGs by introducing timestamps, providing a new perspective for EA. However, while temporal information is critical in TKGs, the EA approach to existing TKG does not fully utilize this critical resource. To solve this problem, we propose a temporal similarity-aware EA model for TKGs (TS-Align). The model innovatively uses the overlap degree of multiple temporal points and temporal intervals of entities to calculate the temporal similarity matrix, and deeply digs and integrates the temporal and relational information in TKGs through a well-designed encoder- decoder architecture. To verify the effectiveness of TS-Align, we conducted experiments on four standard datasets and compared it to other advanced methods. The experimental results show that the accuracy of TS- Align on EA tasks is significantly improved compared with other methods. Specifically, TS-Align is about 0.4 -1.8 % higher on the Hit@1 metric compared to the most advanced time-aware models. This is about 7.4 -17.6 % higher than the most advanced GCN-based models. This is about 20.8 -42.3 % higher than the most advanced translation-based models. In addition, through ablation experiments, we verify that the proposed temporal similarity matrix, relational embedding, and entity embedding contribute to the significant improvement in performance.