.png){kind=icon}
.png){kind=icon}
Return
Continuous-time transformer with large language model for temporal knowledge graph forecasting
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.114695.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Temporal knowledge graph (TKG) forecasting involves predicting future facts or inferring missing information in KGs where relations between entities evolve. A critical method for TKG forecasting is TKG embedding, which aims to learn embedding that captures temporal and structural information. These embedding methods effectively account for both structural and temporal dependencies, facilitating the accurate prediction of future events and the evolution of relationships. However, TKG forecasting presents several challenges: (1) maintaining the continuity of dynamic features in the continuous time domain, (2) addressing the heterogeneity of time representation in TKGs, and (3) achieving collaboration between structural and semantic features. We propose continuous-time Transformer with large language model (CTFormer-LLM) to address these three challenges. The CTFormer-LLM model effectively handles these challenges in a unified framework: (1) we operate a continuous-time Transformer to encode historical events, addressing the continuity of dynamic features in the continuous time domain; (2) we use a Transformer-based approach to learn time embeddings, effectively managing the heterogeneity of time representation; and (3) we combine a LLM and global structural features to improve prediction accuracy, resolving the collaboration between structural and semantic features. Experimental results on benchmark datasets demonstrate that CTFormer-LLM outperforms other models in terms of performance.
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W