Leveraging Relation Attention Mechanisms for Enhanced Knowledge Graph Completion with Embedding Translation

In this paper, we propose a novel knowledge graph completion framework to leverage a relation-specific attention mechanism integrated with an embedding translation strategy to improve the accuracy and contextual understanding of link prediction tasks. Unlike traditional models that rely on fixed transformation spaces, the proposed method dynamically captures fine-grained relational semantics by combining hierarchical candidate categorization, relation-guided entity projection, and asymmetric score functions. Specifically, the proposed model employs K-means clustering and principal component analysis (PCA) to identify semantically consistent entity sets, and integrates attention-weighted multi-attribute embeddings to construct robust relational representations. A margin-based ranking loss with normalized embedding constraints ensures effective optimization, further supported by Xavier initialization and stochastic gradient descent. Extensive experiments on two benchmark datasets, WN18 and FB15K, demonstrate the superiority of the proposed method. Specifically, on WN18, the proposed method achieves the lowest mean rank (MR) of 144, with competitive results in mean reciprocal rank (MRR) (0.902), Hits@1 (89.0%), Hits@3 (90.4%), and Hits@10 (96.3%), closely rivaling state- of-the-art models like QuatE and ComplEx. On FB15K, the proposed method again delivers the best Mean Rank of 21, along with strong scores in MRR (0.831), Hits@1 (72.2%), Hits@3 (88.4%), and the highest Hits@10 (92.5%) among all compared methods.