AGKE: Improve Knowledge Graph Embedding by Adaptively Generating Kernel

Knowledge graph embedding is an important part of knowledge graph research field, and it is the basis of knowledge acquisition and downstream tasks. It enables efficient representation of knowledge by mapping entities and relations into low-dimensional vector spaces. In recent studies, neural network methods for encoding semantic matching have achieved remarkable predictive performance. It brings entities and relations into deep neural networks and computes semantic matching scores. However, some existing methods have weak feature interactions, complex model structures, low generalization ability, and are effective only for specific datasets. Therefore, we propose an improved method for knowledge graph embedding. The model enhances feature interactions by adding an involution module, which adaptively generates convolution kernel parameters based on the input information. The model structure is simple and easy to implement. Our approach is evaluated on several public benchmark datasets. Its evaluation metrics achieve better performance compared to some state-of-the-art models, and the results prove that it is effective.