An Enhanced Knowledge Graph Embedding for Small-Scale Sparse Knowledge Graph

Yushun Xie, Haiyan Wang, Runnan Tan, Xiangyu Song, Zhaoquan Gu

Abstract: Knowledge graph embedding (KGE) learns to map entities and relationships to a continuously dense, high-dimensional representation in a vector space, capturing and utilizing semantic relationships between entities through numerical operations. As the key for realizing intelligent inference for deep learning, it has garnered widespread attention in recent years. However, existing KGE techniques generally perform poorly on small-scale datasets, because small-scale knowledge graphs are easy to cause model overfitting due to their incompleteness and sparsity. To address these issues, in this paper, we propose an enhanced knowledge graph embedding for small-scale knowledge graphs that can extremely efficiently achieve mutual improvement of rule-based data augmentation and neighborhood-based embedding enhancement. Additionally, to make the method applicable to the conducted Cybersecurity Attack Knowledge Graph, we abstract temporal attributes into knowledge, effectively preserving temporal dependencies of the attack steps. Furthermore, we correct the KGE ranking by leveraging spatial rule scores. Extensive experiments show that our method outperforms existing KGE techniques with 15% improvement on MRR and 6% improvement on Hits@10.

External IDs:dblp:conf/dasfaa/XieWTSG25