RotatSAGE: A Scalable Knowledge Graph Embedding Model Based on Translation Assumptions and Graph Neural Networks

Knowledge graphs have been widely used in numerous AI applications. In this paper, we propose an efficient knowledge graph embedding model called RotatSAGE by combining the RotatE model and the GraphSAGE model. In the proposed model the RotatE model is used to learn the embedding vectors of heterogeneous entities and relations in a knowledge graph. One problem of the RotatE model is that it only can learn from a single triplet and cannot take advantage of local information to learn embeddings. To solve this issue, we introduce the GraphSAGE model into RotatE. The GraphSAGE model can use neighbor information to improve the embedding of an entity by sampling a small and fixed number of neighbors. We also propose a sampling strategy to further eliminate redundant entity information and simplify the proposed model. In the experiments, the link prediction task is used to evaluate the performance of embedding models. The experiments on four benchmark datasets show the overall performance of RotatSAGE is higher than baseline models.