Lina Ge, Yongguo Yang, Qinglin Bi, Qinchun Su, Lei Tian, Chaoliang Zhou · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840616
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Continual knowledge graph embedding (CKGE) aims to incrementally learn emerging knowledge in evolving knowledge graphs while preserving previously acquired knowledge. However, existing CKGE approaches suffer from two major limitations. First, they largely overlook the explicit graph structure that is crucial for robust representation learning. Second, newly arrived triples are often learned in a stochastic order, which disrupts topological coherence and fails to preserve the structural dependencies between historical and emerging knowledge, thereby leading to severe catastrophic forgetting. We propose CHR-KGE, a structure-first framework tailored for CKGE under evolving graph scenarios. Specifically, CHR-KGE organizes newly arrived triples into a structure-guided learning order through community partition and hierarchical prioritization, while preserving cross-snapshot structural continuity via structure-constrained reconstruction and aggregation module. Meanwhile, it selectively constrains representation drift in structurally important regions through learnable mask-guided distillation, enabling effective adaptation to new knowledge without sacrificing the stability of historical representations. Experiments on six datasets show that CHR-KGE consistently outperforms strong baselines in link prediction accuracy, knowledge transferability, and training efficiency. Compared with IncDE, CHR-KGE achieves average relative improvements of 3.39% in MRR, 5.36% in Hits@1, and 2.39% in Hits@10. Compared with re-train, CHR-KGE maintains highly competitive accuracy while reducing training time by 40%–60%. These results demonstrate the effectiveness and efficiency of CHR-KGE for continual knowledge graph embedding. The code can be found at https://anonymous.4open.science/r/CHR-KGE/.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.