API documentation update for train_test_split()

Applied train_test_split on the imbalanced dataset (Credit Card Fraud dataset) without setting stratify parameter (None by default). When we checked the test and train data, the class distribution is maintained, i.e., stratification is applied.

Though stratification is applied by default, the document says following which is confusing for users:

stratify : array-like, default=None
    If not None, data is split in a stratified fashion, using this as
    the class labels.
    Read more in the :ref:`User Guide <stratification>`.

API documentation of train_test_split should be updated to reflect the exact behaviour of stratify

[NicolasHug]

If shuffle=false, then none is the only option, no stratification applied.

Yes

is it good to say on the behavior, that stratification is applied if it is specified, given shuffle=true

Yes, if stratify != None then shuffle must be True (and you should probably get an error if you keep it to False)

[NicolasHug]

It’s not a contradiction @brgopalakrishnan , the doc says

shuffle=False => stratify=None

but this doesn’t imply

not(shuffle=False) => not(stratify=None)

A => B is equivalent to not(B) => not(A) but it doesn’t imply not(A) => not(B)

I’ll close the issue since I think I addressed the original issue about imbalance