Abstract Currently, a significant amount of research on Knowledge Graph Embeddings (KGE) focuses on how to represent the relations between entities in vector spaces. However, research on training methods for KGE is relatively limited. This thesis proposes three training methods to explore how different training regimens affect the performance of KGE models. The first method is biased sampling based on the degrees of entities and relations in the knowledge graph (KG), which can be specifically divided into ascending sampling and descending sampling. Experimental results show that this method affects the efficiency and effectiveness of different models across various datasets differently. Only on FB15k-237 is a relatively consistent pattern observed: the ascending sampling slightly improves model performance and converges faster. In contrast, the descending sampling does not improve the performance and even slows down convergence. The second method is the curriculum learning method. Based on the degrees of entities and relations in the KG, the triples in the training set are divided into three groups: easy, medium, and hard. During the training phase, these three groups of triples are introduced sequentially, from easy to hard. Experimental results show that on FB15k-237 and YAGO3-10, the performance of all three models decreases, accompanied by a reduction in training efficiency. \par The third method is the adaptive learning-based training method. This method consists of two phases: initial training with random sampling, followed by fine-tuning using biased sampling based on the results of the first phase. During fine-tuning, the sampling probabilities of triples that were poorly learned in the initial phase, are increased. Experimental results indicate that on FB15k-237, this method improve the performance of RotatE, while it does not have a positive effect on TransE and ComplEx. In Conclusion, the biased sampling and adaptive learning-based training method proposed in this thesis can improve the performance of certain models on specific datasets. \par Key words: Knowledge graph embedding, training regimen, biased sampling, curriculum learning, adaptive learning.