CTHTC: A Hybrid Architecture for Temporal Knowledge Graph Completion
Xinyuan Chen, Mohd Nizam Husen, Zhongmei Zhou · 2024
Temporal knowledge graph completion (TKGC) is of utmost importance for downstream applications. However, most TKGC models fail to capture local interactions and global dependencies simultaneously with evolutionary dynamics, thereby impairing their performance. Latest achievements in convolutions and Transformers have displayed strong capabilities in the computer vision (CV) area but have yet to be employed in TKGC. Furthermore, periodic patterns in temporal knowledge graphs (TKGs) are seldom discussed. To address these issues, a multistage hybrid architecture of convolution-backed Transformers is introduced, combining the Hawkes process and the seasonal-trend decomposition, to model evolving event sequences in a continuous-time domain, and to identify periodic patterns, respectively. Experiments on benchmark datasets are conducted to compare our model against state-of-the-art (SOTA) methods, proving the proposed hybrid architecture's superiority. An extensive ablation study is also carried out to evaluate architectural variants and verify positive contributions brought by components, including the Hawkes process and the seasonal-trend decomposition, etc.