CompGT: a knowledge graph embedding model integrating CompGAT and Comphormer

The existing commonly used knowledge graph completion models exhibit significant deficiencies in capturing information about distant neighboring entities, which severely impacts the effectiveness of knowledge graph completion. To address this issue, this paper designs and proposes a knowledge graph embedding model that integrates Graph Neural Networks (GNN) and Transformer — CompGT. The CompGT model consists of a GNN-based extraction module and a Transformerbased global perception embedding module. The GNN module is composed of CompGAT layers, while the Transformer module consists of Comphormer layers. CompGAT enhances the model’s ability to capture local structural information of entities by incorporating degree structure encoding and attention-based combinatorial operators. Comphormer improves the model’s ability to capture global knowledge graph information by introducing multi-relational fusion degree encoding, distance-based combinatorial encoding, and edge feature encoding. By integrating these two modules, the model effectively improves its capacity to capture both entity structural information and distant neighboring entity information. Experimental results on the public datasets FB15k-237 and WN18RR demonstrate that, compared to baseline models, the proposed model achieves significant advantages across multiple evaluation metrics.