Geometric Knowledge Distillation: Topology Compression for Graph Neural Networks