412604be30f701b1b1e3124c252065e6-AuthorFeedback.pdf

We thankR4 for rightly pointing out that local subgraphs can alleviate over-smoothing50 because, ineach iteration, different subgraphs arefedintoGNN, which promotes inductivegeneralization.(5.4)R451 raised a question about hyper-parameters. We use random search on validation set to select hyper-parameters and52 findthatmodel performance isstable forabroad range ofvalues. Wefollowstandardepisode54 training and semi-supervised setting inwhich most nodes are not labeled, i.e., few-shot learning.