Closed tdhock closed 8 months ago
min_train_data could be interpreted as the min number of samples for the least frequent stratum, and then we would take other strata proportionally. for example we could add a test for the following. 10% class A, 90% class B, 2000 samples overall in train. smallest size would be 10 A / 90 B, largest would be 200 A, 1800 B.
strata are already supported when deciding folds, but not when choosing the random order to go through the train data. would be especially useful for multi-class problems. would be nice to have each sample be a max likelihood according to the multinomial.