Rethinking Efficiency and Redundancy in Training Large-scale Graphs