Simple Negative Sampling for Link Prediction in Knowledge Graphs

Abstract

Knowledge graph (KG) embedding methods learn the low dimensional vector representations of entities and relations of a knowledge graph, facilitating the link prediction task in knowledge graphs. During learning of embeddings, sampling negative triples is important because KGs have only observed positive triples. To the best of our knowledge, uniform-random, generative adversarial network (GAN)-based, and NSCaching, structure aware negative sampling (SANS) are four negative sampling methods in the literature. Unfortunately, they suffer from computational and memory inefficiency problems. In addition, their prediction performance are affected by the ‘vanishing gradient’ problem because of poor quality of sampled negative triples. In this paper, we propose a simple negative sampling (SNS) method based on the assumption that the entities which are closer in the embedding space to the corrupted entity are able to provide high-quality negative triples. Furthermore SNS has a good exploitation potential as it uses sampled high-quality negatives for improving the quality of negative triples in next steps. We evaluate our sampling method through link prediction task on five well-known knowledge graph datasets, WN18, WN18RR, FB15K, FB15K-237, YAGO3-10. The method is also evaluated on a new biological KG dataset (FIGHT-HF-23R). Experimental results show that the SNS improves the prediction performance of KG embedding models, and outperforms the existing sampling methods.