DCNS: A Double-Cache Negative Sampling Method for Improving Knowledge Graph Embedding

Negative sampling plays an important role in knowledge graph embedding. A high-quality negative sample can push the model training to the limit. Most negative triples generated by simple uniform sampling are low-quality negative samples, which will lead to the problem of vanishing gradients in the training process. Generative Adversarial Network (GAN) has been used in the study of negative sampling methods. However, the training for the GAN-based negative sampling method is more complicated. To solve these issues, we propose DCNS. DCNS designs two caches containing high-quality negative triples, samples from the cache and updates the cache. In addition, in order to generate harder negative samples that have a greater impact on training, DCNS adopts a mixing operation. Finally, we evaluated the results of the link prediction model using DCNS on four standard datasets. The extensive experiments show that our method can gain significant improvement on various KG embedding models, and outperform the state-of-the-art negative sampling methods.