Dynamic Knowledge Graph Alignment

Knowledge graph (KG for short) alignment aims at building a complete KG by linking the shared entities across complementary KGs. Existing approaches assume that KGs are static, despite fact almost every evolves over time. In this paper, we introduce task of dynamic knowledge alignment, main challenge which is how to efficiently update entity embeddings evolving topology. Our key insight view parameter matrix GCN as feature transformation operator and decouple process from aggregation process. Based on that, first propose novel base algorithm (DINGAL-B) with topology-invariant mask gate highway gate, consistently outperforms 14 existing methods in static setting. More importantly, it naturally leads two effective efficient algorithms align graph, including (1) DINGAL-O leverages previous matrices affected entities; (2) DINGAL-U resorts newly obtained anchor links fine-tune matrices. Compared their counterpart (DINGAL-B), 10× 100× faster respectively, little accuracy loss.