Abstract: The original full-batch training for dynamic graph neural networks requires calculating the representation of all the nodes in the graph, which is difficult to handle large-scale data due to the limited memory capacity and computing power of current hardware devices. This paper proposes a representation learning strategies of dynamic graph neural network based on layer dependent importance sampling. Some nodes in the dynamic graph were selected as initial nodes, and these initial nodes sampled neighboring nodes layer by layer based on importance probability. The sampled nodes contained the spatiotemporal neighborhood information of the dynamic graph, which could guarantee the model training effect. Experiments were conducted on four publicly available datasets using three different models. The results show that using this strategy for training can reduce the memory overhead of the hardware device and improve the training effect of the model to a certain extent.