Positionally restricted masked knowledge graph completion via multi-head mutual attention

Abstract

Knowledge graph completion aims to enhance the completeness of knowledge graphs by predicting missing links. Link prediction is a common approach for this task, but existing methods, particularly those based on similarity computation, are often computationally expensive, especially for large models. To address this, we propose a novel method, positionally restricted masked knowledge graph completion (PR-MKGC), which reduces inference time by leveraging masked prediction and relying solely on structural information from the knowledge graph, without using textual data. We introduce a multi-head mutual attention mechanism that aggregates neighbor information more effectively, improving the model's ability to predict missing links. Experimental results demonstrate that PR-MKGC outperforms existing models in terms of both predictive performance and inference time on the FB15K-237 dataset.