Simple but Effective Compound Geometric Operations for Temporal Knowledge Graph Completion

TL;DR

The paper addresses temporal knowledge graph completion by introducing TCompoundE, a TKGE model that uses time-specific and relation-specific compound geometric operations to capture temporal dynamics. The head embedding is transformed by time-augmented relation operators and scored against the tail embedding via semantic similarity, trained with reciprocal learning and a temporal regularizer. The authors prove TCompoundE can model symmetric, asymmetric, inverse, and temporal-evolution relation patterns and demonstrate state-of-the-art or competitive results on ICEWS14, ICEWS05-15, and GDELT, with code released. This approach advances TKGE by integrating temporal information into a compact operator framework, offering improved expressivity and practical performance for time-aware link prediction.

Abstract

Temporal knowledge graph completion aims to infer the missing facts in temporal knowledge graphs. Current approaches usually embed factual knowledge into continuous vector space and apply geometric operations to learn potential patterns in temporal knowledge graphs. However, these methods only adopt a single operation, which may have limitations in capturing the complex temporal dynamics present in temporal knowledge graphs. Therefore, we propose a simple but effective method, i.e. TCompoundE, which is specially designed with two geometric operations, including time-specific and relation-specific operations. We provide mathematical proofs to demonstrate the ability of TCompoundE to encode various relation patterns. Experimental results show that our proposed model significantly outperforms existing temporal knowledge graph embedding models. Our code is available at https://github.com/nk-ruiying/TCompoundE.

Figures (4)

• Figure 1: An illustration of temporal evolution patterns. It can be observed that the relationships between head entity and tail entity are dynamically determined by both relation and time.
• Figure 2: An illustration of translation, rotation and scaling. Where $\bm{v}$ is the original vector, $\bm{T \cdot v}$ is the vector after the translation operation, $\bm{R \cdot v}$ is the vector after the rotation operation, and $\bm{S \cdot v}$ is the vector after scaling.
• Figure 3: Visualisations of the learned entity embeddings on ICEWS14. The circle in the figure represents the head entity and the triangle represents the tail entity; Circles and triangles of the same color indicate that the head and tail entities come from the same quadruple.
• Figure 4: MRR scores on ICEWS14 dataset.

Theorems & Definitions (8)

• Proposition 1
• Proposition 2
• Proposition 3
• Proposition 4
• Definition 1
• Definition 2
• Definition 3
• Definition 4