Scholarlink

<p>Temporal fact prediction in knowledge graphs is a crucial task with widespread applications, including recommendation systems and natural language understanding. Existing methods, such as tensor factorization and neural networks, often suffer from high computational costs and lack transparency due to the complex nature of temporal data. In this work, we introduce AttBiTi (Attention-enhanced Bicomplex Temporal Integration), a novel model that leverages bicomplex embeddings with implicit attention to improve the representation of entities, relations, and time in temporal knowledge graphs (TKGs). Unlike traditional models that rely on explicit attention mechanisms, AttBiTi employs implicit attention-driven temporal translation to adaptively weight relational and temporal influences. By integrating adaptive bicomplex transformations, AttBiTi effectively captures evolving temporal dynamics while preserving algebraic transparency. The model utilizes translation to refine temporal dependencies and bicompound rotations to model complex relational structures. Experimental results demonstrate that AttBiTi achieves a 21% improvement in Mean Reciprocal Rank (MRR) and Hits@k (Hits at K) compared to state-of-the-art methods, as evaluated on benchmark datasets, including ICEWS (Integrated Crisis Early Warning System), GDELT (Global Database of Events, Language, and Tone), and YAGO11k (Yet Another Great Ontology). Furthermore, ablation studies highlight the significance of temporal-aware attention mechanisms, as removing temporal encoding leads to a notable drop in MRR from 84.4 to 78.5 on ICEWS14, reinforcing the necessity of adaptive time integration in TKGs.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Temporal-Aware bicomplex embeddings with implicit attention for knowledge graph link prediction

  • 期刊论文

  • 发表日期：

    2025年10月31日

查看全文
  • Thanh Le,
  • Trong-Nghia Pham

摘要

Temporal fact prediction in knowledge graphs is a crucial task with widespread applications, including recommendation systems and natural language understanding. Existing methods, such as tensor factorization and neural networks, often suffer from high computational costs and lack transparency due to the complex nature of temporal data. In this work, we introduce AttBiTi (Attention-enhanced Bicomplex Temporal Integration), a novel model that leverages bicomplex embeddings with implicit attention to improve the representation of entities, relations, and time in temporal knowledge graphs (TKGs). Unlike traditional models that rely on explicit attention mechanisms, AttBiTi employs implicit attention-driven temporal translation to adaptively weight relational and temporal influences. By integrating adaptive bicomplex transformations, AttBiTi effectively captures evolving temporal dynamics while preserving algebraic transparency. The model utilizes translation to refine temporal dependencies and bicompound rotations to model complex relational structures. Experimental results demonstrate that AttBiTi achieves a 21% improvement in Mean Reciprocal Rank (MRR) and Hits@k (Hits at K) compared to state-of-the-art methods, as evaluated on benchmark datasets, including ICEWS (Integrated Crisis Early Warning System), GDELT (Global Database of Events, Language, and Tone), and YAGO11k (Yet Another Great Ontology). Furthermore, ablation studies highlight the significance of temporal-aware attention mechanisms, as removing temporal encoding leads to a notable drop in MRR from 84.4 to 78.5 on ICEWS14, reinforcing the necessity of adaptive time integration in TKGs.