more group-based splitting methods

[topepo]

It would be good to have an initial_group_split(data, group, strata, prop) method that can split the data when there are groups (perhaps patients). The strata option might be difficult when the outcome (or other stratification variable) is not constant within each group. We could also use the median or mode on the stratification variable and use that.

Similarly, a mc_group_cv() function would also be a good idea (using the splitting method as above).

[NateMietk]

I second this feature! I have pre-defined testing/training datasets so I cannot use the split_initial function, which in turn does not allow me to utilize the the last_fit function. This would open up the world of tidymodels for me and other uses like me!

[juliasilge]

@NateMietk It sounds like you may be interested in a different kind of feature, one that we already support. If you want to create an initial split from your own training/testing data, then you want to use rsample::make_splits(). Check out:

• this Stack Overflow answer
• this GitHub issue

for some examples on how to use it.

[hnagaty]

I second this too. I came across a data set that has a grouping variable. I already used group_vfold_cv() for the cross validation resamples, and it would be helpful to have a similar approach to the initial split.

[rkb965]

This would be a great idea. :) I'm trying to bootstrap with repeated measures data and want all measurements from an individual to be in the same split. It seems like this grouped initial_split would enable this, correct? Do y'all have any ideas for work arounds in the meantime? Thanks for all of the incredible work on tidymodels!

[juliasilge]

Yes @rkb965 in the meantime, you can create a custom split with make_splits(). Check out:

• this Stack Overflow answer
• this GitHub issue

for some examples on how to use it.

[astamm]

Assuming that stratification variable stratifies the groups, could something like this have any sense:

group_initial_split <- function(data, group, prop = 3/4, strata = NULL, breaks = 4, pool = 0.1, ...) {
  strata <- enquo(strata)
  group <- enquo(group)
  data <- nest(data, data = !any_of(c(!!group, !!strata)))
  cell_split <- initial_split(data, prop, !!strata, breaks, pool, ...)
  train_id <- data$id[cell_split$in_id]
  cell_split$data <- unnest(cell_split$data, cols = data)
  cell_split$in_id <-  which(cell_split$data$id %in% train_id)
  cell_split
}

[juliasilge]

We'll want to add group as an argument to mc_cv(), initial_split(), and vfold_cv(). (Nothing else, right?)

Since initial_split() uses mc_cv() internally, let's make the changes in mc_cv(), specifically mc_splits() (along with vfold_splits()).

[MSHelm]

Having the group argument for validation_split() would also be very helpful for my case. Then I could directly use it for hyperparameter tuning, using e.g. tune_grid

[mattwarkentin]

Related: #284

[mikemahoney218]

From conversation with Julia (where we talked about this in reference to spatialsample), it seems like adding _group_ functions would be better than adding group arguments. You can't specify both group and strata (plus pool and breaks), which introduces some nasty dependencies between arguments. Separate functions avoids the issue.