Scholarlink

<p>With the rapid advancement of the Internet and artificial intelligence technologies, knowledge graphs have demonstrated significant potential in data structuring and knowledge reasoning. In particular, temporal knowledge graphs (TKGs), have attracted increasing attention for their applications in dynamic scenarios. However, traditional TKG representation learning methods often overlook the topological structure among entities and their interactions with neighboring nodes, resulting in limited capacity to model complex historical dependencies and perform global reasoning. To address these challenges, we propose TG-CENET (Transformer‑GNN‑Contrastive Event Network), an enhanced temporal knowledge graph reasoning model that builds upon the historical contrastive learning strategy of CENET. By incorporating a GNN and a Transformer module, TG-CENET improves both structural representation and temporal sequence modeling. Specifically, the GNN component captures structural information from the graph, improving the quality of entity representations by modeling neighborhood relationships. Meanwhile, the Transformer module captures long-term temporal dependencies, enabling the model to reason more effectively over both historical and non-historical events. The effectiveness and performance of the proposed method were evaluated in four representative public datasets, ICEWS14s, ICEWS18, GDELT and YAGO. The results demonstrate that TG-CENET consistently outperforms existing methods in multiple metrics, including Hit@1, Hit@3, Hit@10 and MRR. The proposed method achieves Hit@1, Hit@3, Hit@10 and MRR scores of 0.31, 0.46, 0.62 and 0.41 on the ICEWS14s, 0.19, 0.32, 0.48 and 0.28 on the ICEWS18, 0.14, 0.22, 0.38 and 0.21 on the GDELT, and 0.49, 0.70, 0.83 and 0.61 on the YAGO, respectively. Ablation studies further confirm the complementary benefits of GNN-based structural encoding, Transformer-based temporal reasoning, and contrastive learning. These results highlight the effectiveness and robustness of TG-CENET in modeling complex temporal dynamics, offering a promising direction for advancing TKG reasoning in real-world applications such as event forecasting, financial analysis, and urban disaster management.</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

TG-CENET: an improved reasoning model for temporal knowledge graphs based on contrastive history

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月17日

查看全文
  • Lizhi Miao,
  • Kaiwen Wu,
  • Yi Huang

摘要

With the rapid advancement of the Internet and artificial intelligence technologies, knowledge graphs have demonstrated significant potential in data structuring and knowledge reasoning. In particular, temporal knowledge graphs (TKGs), have attracted increasing attention for their applications in dynamic scenarios. However, traditional TKG representation learning methods often overlook the topological structure among entities and their interactions with neighboring nodes, resulting in limited capacity to model complex historical dependencies and perform global reasoning. To address these challenges, we propose TG-CENET (Transformer‑GNN‑Contrastive Event Network), an enhanced temporal knowledge graph reasoning model that builds upon the historical contrastive learning strategy of CENET. By incorporating a GNN and a Transformer module, TG-CENET improves both structural representation and temporal sequence modeling. Specifically, the GNN component captures structural information from the graph, improving the quality of entity representations by modeling neighborhood relationships. Meanwhile, the Transformer module captures long-term temporal dependencies, enabling the model to reason more effectively over both historical and non-historical events. The effectiveness and performance of the proposed method were evaluated in four representative public datasets, ICEWS14s, ICEWS18, GDELT and YAGO. The results demonstrate that TG-CENET consistently outperforms existing methods in multiple metrics, including Hit@1, Hit@3, Hit@10 and MRR. The proposed method achieves Hit@1, Hit@3, Hit@10 and MRR scores of 0.31, 0.46, 0.62 and 0.41 on the ICEWS14s, 0.19, 0.32, 0.48 and 0.28 on the ICEWS18, 0.14, 0.22, 0.38 and 0.21 on the GDELT, and 0.49, 0.70, 0.83 and 0.61 on the YAGO, respectively. Ablation studies further confirm the complementary benefits of GNN-based structural encoding, Transformer-based temporal reasoning, and contrastive learning. These results highlight the effectiveness and robustness of TG-CENET in modeling complex temporal dynamics, offering a promising direction for advancing TKG reasoning in real-world applications such as event forecasting, financial analysis, and urban disaster management.