LightHGNN: Distilling Hypergraph Neural Networks into MLPs for $100\times$ Faster Inference