Efficient Training of Graph Neural Networks for Dynamic Phenomena (Proposal)