e347c51419ffb23ca3fd5050202f9c3d-AuthorFeedback.pdf

Reviewer 1: "The algorithm is rather similar to the denoising algorithm referred to in the paper. One could argue that the prediction time17 iswhat really matters. The preprocessing time (including tuning) is at the order of 4,00028 seconds, compared tohundred seconds using ourmethod. Interms ofprediction time, thebest speedup (x8) comes29 at the cost of the worst accuracy(2 times our regret) while the speedup with their best accuracyis only 2to 3folds30 (compared to10folds inourmethods). Perhaps some discussion onunderstanding how much49 training data we could get away with (and whether we could just ignore a lot of the data to save on computation)50 could behelpful..." Indeed, evenXueandKpotufe (2018) suggested thatasmall number ofsubsamples (repititions)51 wouldsuffice.