Temporal rotation knowledge graph completion based on hybrid attention mechanism
Wang Lulu
School of Information Engineering, Dalian University
Abstract： A novel model that integrates a hybrid attention mechanism with temporal rotation is proposed to address the limitations of capturing dynamic relation patterns, handling asymmetric, temporary, and reflexive relations in existing temporal knowledge graph completion. On one hand, by introducing temporal rotation, we leverage vectors in complex spaces to represent entities and relations evolving over time, especially to handle relation changes within temporal intervals. The adoption of a dual-complex embedding scheme significantly enhances the expressive power for temporal characteristics. On the other hand, by introducing spatial attention and channel attention to analyze the knowledge graph from two dimensions, the model can better focus on the most crucial entity and relation features in the temporal sequence for prediction, thus mining temporal correlation information from complex time series. Through experimental evaluations on the ICEWS14, ICEWS18, YAGO11k, and WIKI12k datasets, the model outperforms baseline models in terms of MRR, Hits@1, Hits@3, and Hits@10, demonstrating the superiority and strong robustness of the proposed algorithm.
Key words : temporal knowledge graph; temporal rotation; hybrid attention mechanism; link prediction
