A Continual Knowledge Graph Embedding Method Based on Local-Global Distillation

Knowledge Graph Embedding (KGE) technology, which maps entities and relations into low-dimensional vector spaces, has been widely applied in tasks such as question answering, semantic search, and link prediction. However, traditional KGE models often assume that knowledge graphs are static and cannot easily adapt to the dynamic nature of knowledge graphs that frequently update in real-world. This paper proposes a novel Continual Knowledge Graph Embedding (CKGE) method, which introduces a hierarchical backpropagation mechanism and a local-global distillation module to fully leverage the topological structure features of the graph and effectively mitigate the catastrophic forgetting problem. Through experiments on four benchmark datasets, the performance advantages of the proposed method are verified. Compared to existing methods, this method performs better in terms of dynamic adaptability of knowledge graphs and retention of old knowledge, demonstrating its efficiency and effectiveness in continual learning of knowledge graphs.