Knowledge Augmented Attention Gating Embedding for Link Prediction

Knowledge graphs (KGs) use graphical representations to illustrate relations between entities, forming a structured knowledge base. However, KGs often suffer from incompleteness, lacking some entities and their interactions. Current link prediction technologies based on KG embeddings primarily focus on learning entity features and relation patterns from existing KGs to address the incompleteness issue. Yet, these methods encounter the inherent sparsity issue in KGs—uneven distribution of entity connections and scarcity of direct links. This complexity challenges models in identifying intricate, potential associations, thereby compromising link prediction accuracy and overall model performance. In this paper, we introduce a innovative Knowledge Augmented Attention-based Gating Embedding (KAAGE) framework for link prediction. Our model augments knowledge by reversing relations and triples, then integrates both original and augmented structural data through gating units. Extensive experiments on three established benchmark datasets demonstrate the superiority of KAAGE over current models. Our implementation code is available at https://github.com/22zwChen/KAAGE