STCKGE: Continual knowledge graph embedding based on spatial transformation

Xinyan Wang, Jinshuo Liu, Kaijian Xie, Meng Wang, Cheng Bi, Juan Deng, Donghong Ji, Jeff Z. Pan

Abstract: Highlights•Proposes STCKGE, a novel spatial transformation-based CKGE framework featuring dual-component entity representations (base vector + offset vector) and relation-aware spatial regions, significantly enhancing complex relational modeling (e.g., multi-hop) while reducing new knowledge dependency on historical embeddings.•Introduces a Bidirectional Collaborative Update (BCU) strategy that efficiently propagates knowledge through lightweight offset vector operations, minimizing retraining costs for both new and historical knowledge.•Experimental results confirm STCKGE’s strong performance in multi-hop relationship learning and prediction accuracy, with an average MRR improvement of 5.4 %.•Constructs and releases the MULTI benchmark dataset with explicit multi-hop facts, addressing selection bias in existing CKGE benchmarks.

External IDs:dblp:journals/kbs/WangLXWBDJP25