Abstract: A knowledge representation learning method based on neighborhood granularity and three-way decision theory(NGTwD)is proposed. The method is implemented using a two-stage enhancement algorithm framework. In the first stage, knowledge representation learning is utilized to fit the nodes and relations in the knowledge graph, and map the embedded semantic information into a low-dimensional vector space. To better capture and exploit the latent similarities in the semantic information, the neighborhood granularity of the low-dimensional vector representations is divided in the second stage. This process is further complemented by applying three-way decision theory to precisely segment the similar semantic information. The extracted latent information is then used to retrain the model, thereby improving the accuracy and robustness of the knowledge representation learning method. Five classic knowledge representation learning models are selected, and experiments are conducted on four large publicly available knowledge graph datasets. The effectiveness of the proposed method is validated through the experimental results.

Key words: knowledge graph, knowledge graph representation learning, neighborhood granularity, three-way decision