Reinforced Negative Sampling for Knowledge Graph Embedding

As a key part of knowledge graph embedding (KGE), negative sampling can mine hard negative examples required for model training to improve the accuracy and effectiveness of the KGE models. However, static negative sampling can’t adapt to the dynamic distribution of hard negative examples, causing the model to miss some important negative examples during the model training. Dynamic negative sampling, although it can adapt to the changes in the data, is unable to mine and select enough hard negative examples. So most existing negative samplings affect the model’s comprehension and representation of the data. In this paper, we propose a negative example sampling method by mixup (NESI), which can generate a large number of synthetic examples at the feature-level to solve the problems of lack of hard negative examples and gradient vanishing. Our method utilizes two data mixing strategies and four evaluation functions to rapidly expand the size of the negative candidates, and effectively improve the quality of them. Therefore, NESI can efficiently select enough high-quality hard negative examples. Compared with the existing negative samplings, NESI improves the model performance on MRR by 10.58% and greatly accelerates the convergence speed of the KGE models.