Abstract
Temporal knowledge graphs (TKGs) represent real-world facts evolving over time by incorporating the multirelational structure of knowledge graphs with temporal dimensions. While extensions of graph convolutional networks, such as relational and composition-based variants, have demonstrated strong capabilities in modeling graph-structured data, effectively capturing both symmetric and asymmetric relational dependencies, their application to TKG reasoning remains challenging due to the need for preserving graph topology while integrating temporal dynamics. In this study, we propose a novel approach that explicitly models the compositional interactions between entities and relations within a complex geometric space, ensuring structural consistency while accommodating diverse relational patterns. To further enhance temporal reasoning, we introduce a diachronic mechanism that adaptively fuses structural graph information with temporal evolution. Through extensive experiments on multiple benchmark datasets using different composition operators, our approach consistently outperforms state-of-the-art methods across most metrics. These results highlight the effectiveness of our approach in capturing both spatial and temporal dependencies for robust TKG reasoning.
Le, T., Loc, X. (2026) Knowl Inf Syst.