Scholarlink

We introduce TempHypE, a novel temporal knowledge graph embedding framework that unites hyperbolic geometry with Neural Ordinary Differential Equations (ODEs) to capture both hierarchical structure and continuous-time evolution. Unlike previous models that relied on Euclidean embeddings or discretized time snapshots, TempHypE embeds entities and relations in the Poincaré ball, enabling temporally smooth and geometrically consistent link prediction. Experiments on dynamic benchmarks ICEWS18 and GDELT show that TempHypE substantially outperforms leading baselines, including TNTComplEx, TA-DistMult, and HyperKG, achieving up to 43% higher mean reciprocal rank (MRR), 37% lower mean average rank (MAR), and 30% higher hits @ 10. Furthermore, TempHypE demonstrates up to 40% lower standard deviation in MAR, indicating greater robustness and stability over time. These advances are driven by TempHypE’s temporal smoothness regularization and Möbius-based hyperbolic operations, supporting fine-grained, interpretable, and consistent predictions. Collectively, these properties make TempHypE a compelling choice for real-world applications that require reliable temporal reasoning in evolving relational structures.

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

TempHypE: Time-Aware Hyperbolic Neural Ordinary Differential Equation (ODEs) Knowledge Graph Embeddings For Dynamic Link Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Amangel Bhullar,
  • Ziad Kobti

摘要

We introduce TempHypE, a novel temporal knowledge graph embedding framework that unites hyperbolic geometry with Neural Ordinary Differential Equations (ODEs) to capture both hierarchical structure and continuous-time evolution. Unlike previous models that relied on Euclidean embeddings or discretized time snapshots, TempHypE embeds entities and relations in the Poincaré ball, enabling temporally smooth and geometrically consistent link prediction. Experiments on dynamic benchmarks ICEWS18 and GDELT show that TempHypE substantially outperforms leading baselines, including TNTComplEx, TA-DistMult, and HyperKG, achieving up to 43% higher mean reciprocal rank (MRR), 37% lower mean average rank (MAR), and 30% higher hits @ 10. Furthermore, TempHypE demonstrates up to 40% lower standard deviation in MAR, indicating greater robustness and stability over time. These advances are driven by TempHypE’s temporal smoothness regularization and Möbius-based hyperbolic operations, supporting fine-grained, interpretable, and consistent predictions. Collectively, these properties make TempHypE a compelling choice for real-world applications that require reliable temporal reasoning in evolving relational structures.