Quaternion knowledge graph embedding with dual bidirectional rotation and restriction relation rotation

Knowledge graph embedding aims to address the inherent incompleteness of real-world knowledge graphs. Although the current quaternion-based model DaBR fuses semantic matching and geometric distance to improve link prediction, it remains consreained by two critical limitions: reliance on single-path of semantic interactions between entities, which makes it difficult to capture complex dependencies; and the linear translation constraint, which leads to insufficient modeling capability for key relation patterns. To overcome these limitations, we propose an improved quaternion-based KGE model with dual bidirectional rotation and restriction relation rotation (DBRR). It contains two major innovations: a dual bidirectional rotation mechanism in the semantic module and a dynamic rotation translation mechanism in the distance module. Specifically, in the quaternion space, to independently capture heterogeneous and complex interactive dependencies between entities, we first introduce two independent rotation vectors for each relation. Then we use these introduced independent rotation vectors to perform a right rotation on the head entity and a reverse rotation on the tail entity, respectively. Subsequently, we perform a right rotation on the transformed head entity and a reverse rotation on the transformed tail entity through the relation itself. In the geometric distance scoring module, we dynamically rotate the head and tail entities to adjust their target directions to facilitate more accurate alignment of the translation terms and enhance the relation representation capability of the model. We also provide mathematical proof to demonstrate that our model can handle complex logical relations. We conducted experiments on five benchmark datasets, and the results and analysis show that our model outperforms previous models on most metrics, confirming the effectiveness of these dual rotation and dynamic translation mechanisms.