Research Reveals Overlooked Entity Interference Problem in Continual Knowledge Graph Embeddings

A new research paper challenges current approaches to Continual Knowledge Graph Embedding (CKGE) methods, revealing a significant oversight in how catastrophic forgetting is assessed. Published on arXiv under identifier 2604.19401v1, the study demonstrates that existing CKGE evaluation protocols systematically overestimate performance by failing to account for entity interference. When new entities are introduced to evolving knowledge graphs, their embeddings can interfere with previously learned ones, causing models to incorrectly predict new entities in place of previously correct answers. This phenomenon, termed entity interference by the researchers, has been largely overlooked despite its substantial impact on model performance degradation. Current CKGE approaches primarily address catastrophic forgetting by limiting changes to existing embeddings, but this perspective proves incomplete according to the findings. Knowledge Graph Embeddings (KGEs) support numerous downstream tasks over Knowledge Graphs (KGs), which in practice evolve as new entities and facts are added. The research indicates that without proper accounting for entity interference, assessments of catastrophic forgetting become misleading. The paper calls for revised evaluation protocols that incorporate this previously neglected factor to provide more accurate performance measurements for CKGE methods.