Scholarlink

<p>Predicting incomplete temporal knowledge graphs (TKGs) presents a significant challenge, as it requires effectively capturing both the structural dependencies between concurrent facts and the dynamic evolution of relationships over time. To address this, we propose the Self-aware and Relation-aware Recurrent Evolution Network (SRREN), a novel recursive algorithm for dynamic representation learning in TKGs. SRREN integrates an evolution unit and a score calculation unit, where the former combines two synergistic modules: (1) The self-aware mechanism employs dual-channel attention with global averaging and local median scaling within fixed time windows, leveraging GRU-based gating to balance long-term trends and outlier-resistant local features. (2) The relation-aware mechanism introduces a GRU-enhanced memory bank and adaptive relation weighting, coupled with GCN propagation to model cross-subgraph relationship evolution. The score unit further incorporates GCN-based decoders to enhance entity-relation interaction modeling. Crucially, SRREN unifies static entity attributes with dynamic relational features through recursive temporal gating, enabling context-sensitive adaptation. Experiments on six benchmark (ICEWS18, GDELT, etc.) show that SRREN can improve the accuracy of predicting future facts, and its performance is better than most other current models.</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

SRREN: self-aware and relation-aware recurrent evolution network for temporal knowledge graph reasoning

  • 期刊论文

  • 发表日期：

    2026年3月13日

查看全文
  • Laibin Zhao,
  • Kuru Ratnavelu,
  • Ghassan Saleh ALDharhani

摘要

Predicting incomplete temporal knowledge graphs (TKGs) presents a significant challenge, as it requires effectively capturing both the structural dependencies between concurrent facts and the dynamic evolution of relationships over time. To address this, we propose the Self-aware and Relation-aware Recurrent Evolution Network (SRREN), a novel recursive algorithm for dynamic representation learning in TKGs. SRREN integrates an evolution unit and a score calculation unit, where the former combines two synergistic modules: (1) The self-aware mechanism employs dual-channel attention with global averaging and local median scaling within fixed time windows, leveraging GRU-based gating to balance long-term trends and outlier-resistant local features. (2) The relation-aware mechanism introduces a GRU-enhanced memory bank and adaptive relation weighting, coupled with GCN propagation to model cross-subgraph relationship evolution. The score unit further incorporates GCN-based decoders to enhance entity-relation interaction modeling. Crucially, SRREN unifies static entity attributes with dynamic relational features through recursive temporal gating, enabling context-sensitive adaptation. Experiments on six benchmark (ICEWS18, GDELT, etc.) show that SRREN can improve the accuracy of predicting future facts, and its performance is better than most other current models.