Dual distillation knowledge embedding framework for efficient knowledge graph completion

Knowledge graph embedding (KGE) models, known for their strong inference capabilities, face challenges in deployment on resource-constrained edge devices. To address this, we propose D²KE, a novel knowledge distillation framework for optimizing high-dimensional KGE models. D²KE employs two teacher models with different dimensions to guide a dual iterative training process, enabling effective compression of high-dimensional embeddings into low-dimensional representations. Additionally, a dynamic soft label weight adjustment strategy is introduced to enhance the student model’s learning efficiency. Experimental results confirm that D²KE achieves superior distillation performance across diverse KGE models.