I realise the values for sample_size in boost_tree represent the proportion (in XGB only, absolute number otherwise) of observations to be subsampled per iteration(tree) but what I don’t understand is how the proportions can be so high (e.g. .49) when sampling is supposed to be random without replacement. A proportion of .49 of the dataset per tree would surely mean repetition across 1000 trees? I don’t know if this is due to my lack of understanding of how subsamples are selected but any help would be greatly appreciated!

I looked at the various parameter combinations in the grid to understand how subsampling works (I’ve used kfold cross validation/resamples and trying to understand how that interacts with subsampling during training). I checked ‘sample_size’ to try to understand how subsamples are selected but found the proportions to be much higher than expected considering boosting uses random sampling without replacement (each observation only used once).