Does the gate work on a donor it was not fitted to

Refits the whole strategy with one donor withheld and scores it on that donor, once per donor. Returns the per-donor scores and their spread.

Usage

gate_transferability(
  X,
  y,
  donor,
  min_donors = 3L,
  max_donors = 8L,
  max_cells = 20000L,
  seed = 42L,
  ...
)

Arguments

X

marker matrix

y

0/1 target indicator

donor

grouping vector, one entry per row of X

min_donors

refuse below this many donors, where the statistic would be an anecdote

max_donors

ceiling on folds

max_cells

ceiling on training rows per fold

seed

seed for the local RNG stream

...

passed to explain_cluster()

Value

list(per_donor = data.frame, summary = data.frame), or NULL

Details

READ THE MINIMUM, NOT THE MEAN. A gate that transfers is one whose worst donor is acceptable. The mean hides exactly the failure this function exists to expose.

COST, AND THE TWO CAPS THAT BOUND IT. Every fold refits the whole strategy, so the work is one fit per donor per label and grows with both. On a cohort of twenty-odd donors and half a dozen labels an uncapped version runs for hours, which is not a validation statistic anyone will wait for.

max_donors bounds the number of folds. The subset is drawn at random from the fixed seed rather than taken as the largest donors, because a worst-donor statistic computed only on the best-represented donors is optimistic in exactly the direction that matters. max_cells subsamples each fold's TRAINING rows, stratified by label; the held-out donor is always scored in full. Both caps are logged when they bind, so a summary over eight donors is never mistaken for one over twenty.