New TeRoR Model Boosts Temporal Knowledge Graph Embeddings

Temporal Knowledge Graphs (TKGs) are crucial for understanding how relationships between entities evolve over time. While existing embedding methods like TeRo have made strides, they often struggle with accurately representing complex relational mapping properties, such as one-to-many or many-to-one relationships, and fully capturing temporal dynamics. A new approach, TeRoR, addresses these limitations by introducing a decoupled temporal rotation mechanism. This allows for independent rotational transformations of head and tail entities in a complex vector space, significantly strengthening the model's ability to capture temporal information. Furthermore, TeRoR incorporates a relational circular region, where a learned radius constrains rotated and translated head entities around the tail entity. This innovative technique effectively models the diverse mapping properties of relations, leading to improved performance on various TKG datasets compared to state-of-the-art models.