Towards Faster Graph Partitioning via Pre-training and Inductive Inference