Periodic Patterns and Long-Term Dependencies Based Temporal Knowledge Graph Completion

Penghui Ge, Wei Chen, Xi Chen, Qingzhi Ma, Lei Zhao

Abstract: Temporal knowledge graph completion (TKGC) under the extrapolation setting aims to predict future facts at unknown timestamps. Although many historical facts exhibit both periodic patterns and long-term dependencies, the existing studies only consider the one of them, i.e., the integration of the two features has not been explored so far. To fill the gap, we propose a dual-graph neural network model PPLD (Periodic Patterns and Long-term Dependencies based Temporal Knowledge Graph Completion). Specifically, the model first explores the periodic patterns across adjacent timestamps from the constructed periodic-aware graphs. Then, the long-term dependencies among historical entities are extracted on a global background graph. Next, the representations of periodic patterns and long-term dependencies are integrated into a unified one through a gating structure. Finally, we introduce a contrastive learning-based module to mitigate the interference from intricate similar paths stemming from multiple historical entities. The extensive experiments conducted on the ICEWS dataset substantiate the efficacy of our proposed model.