Knowledge graph embedding by relational rotation and complex convolution for link prediction

Knowledge graphs are organized as triplets to represent facts from the real world and play an important role in various intelligent information systems. Because knowledge graphs are frequently constructed using manual or semi-automatic methods, they often miss connections between entities. Link prediction was created to solve this problem. Many recent state-of-the-art studies, such as those introducing the RotatE and RotatHS models, have advocated for rotation transformations with entity and relation embeddings in complex vector spaces. However, using only rotation planes means that these models do not have the expressive power of models based on neural networks, such as the ConvE and the ConvR models. As a result, link prediction performance suffers. To address these shortcomings, this paper proposes the ConvRot model, which integrates a 2D convolution. Specifically, we perform convolution on embeddings of entities and relations to obtain support vector embeddings. These vectors are then integrated into an element-wise rotation from the head entity to the tail entity using the Hadamard product, enabling the model to capture local interactions among entities and relations through the neural network while still ensuring intuitiveness through a roto-transformation in the link prediction. In addition, we present two strategies for designing the complex convolution module and show their effects on model performance. The proposed method is evaluated on standard benchmark datasets and achieves significantly improved results on MRR and [email protected]K (K = 1, 3, 10). Overall, our model’s link prediction performance is superior by approximately 5–7 %. Moreover, the ConvRot model is also considered separately on many relation types, such as one-to-one, one-to-many, many-to-one, and many-to-many. Finally, we prove that type constraints can help increase the model’s overall performance, especially on complex and large datasets.