Jointly leveraging 1D and 2D convolution on diachronic entity embedding for temporal knowledge graph completion

Abstract

Temporal knowledge graphs (TKGs) model knowledge that dynamically changes over time in the real world, providing effective support for temporal-aware artificial intelligence (AI) applications. However, existing TKGs are far from complete, and their incompleteness significantly affects the performance of downstream applications. Therefore, Temporal Knowledge Graph Completion (TKGC) has become a current research hotspot, which aims to reason potential missing facts based on existing ones. In the widely studied TKGC methods with the implicit representation of temporal information, existing methods that embed temporal information into entity representations can capture the temporal evolution of entities. However, they fail to take the behavioral characteristics of entities across different time units into account, making them challenging to precisely model the fine-grained dynamics of entities. Furthermore, given the powerful expressiveness of Convolutional Neural Networks (CNNs), some TKGC methods have employed the 1D convolution operation to capture global relationships within the embedded quadruple, enabling the learning of explicit knowledge in TKGs and attaining competitive performance for TKGC. Nevertheless, the non-linear and deep features embedded in the entity-relation interaction have not been insufficiently explored. To address these challenges, this paper proposes JointDE, a TKGC model that applies both 1D and 2D convolution operations to the generated diachronic entity embedding, which simultaneously learns the explicit and implicit knowledge in TKGs. The new diachronic entity embedding method explicitly models the inherent attributes of entities and integrates temporal features across different time units, thereby possessing the ability to capture fine-grained entity evolution. More importantly, we construct feature matrices and filters using diachronic entity embeddings and relation embeddings, leveraging an internal 2D convolution mechanism to expand their interactions. This is the first work to learn implicit knowledge embedded in TKGs from a local relationship perspective for TKGC. Experimental results demonstrate that JointDE surpasses several TKGC baseline methods and achieves state-of-the-art performance on three event-based benchmark datasets: ICEWS14, ICEWS05–15, and GDELT. Specifically, JointDE improves Mean Reciprocal Rank (MRR) by 3.17 % and Hits@1 by 5.87 % over the state-of-the-art baseline for entity reasoning.