Xudong Wang, Xiaotian Yu, Dan Li, See-Kiong Ng, Bo Zhang · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840814
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).
Real-world knowledge graphs evolve continuously, making retraining from scratch impractical. Continual knowledge graph embedding (KGE) updates representations with new triples while preserving historical knowledge. However, current methods face two challenges: (1) the stability-plasticity dilemma is pronounced for bridge entities connecting old and new subgraphs, where uniform updates may cause forgetting or rigidity; and (2) random negative sampling yields easy negatives and weak supervision near evolving graph boundaries. We propose KG-BRACE, a bridge-adaptive continual KGE framework combining topology-aware negative sampling with differentiated low-rank updates. KG-BRACE learns new entities via DoRA, updates bridge entities through a restricted Mini-DoRA branch, and freezes inactive historical entities. It also retrieves hard negative candidates from local neighborhoods. Experiments on six benchmarks show best or tied-best results on five datasets, consistent Hits@1 improvements, and competitive stronger freezing in extremely sparse graphs.
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.