Learning Representations of Tail Entities via Interactions and Scalar Fusions Among Subspaces of Head Entities and Relationships

In the past few decades, algorithms for chain prediction and knowledge graph completion tasks have been extensively studied. However, traditional methods based on geometric spatial embedding face challenges such as high dimensionality, high computational complexity, and poor scalability. To address these issues, we propose a novel model: Representations of Tail Entities from the Interactions and Scalar Fusions be-tween the Subspaces of Head Entities and Relationships (RTIFS). This model is based on a new assumption of the spatial relationship between head-relation-tail triples, embedding and abstracting tail entity information from the combined space. Experiments show that our model achieves competitive results compared to the latest eight state-of-the-art baselines, with an average reduction in embedding dimensions of 68.8%.