flip_y in make_classification is misleading

As per description of flip_y, it is “The fraction of samples whose class are randomly exchanged.” So when you have two classes one would expect by setting flip_y equal to 0.1, 10% of the labels flip (exchange), as the name suggest (flip_y). However, if you look at the source code 10% of the labels are assigned random labels which 50% of the time they are assigned their own labels so about 5% of labels are going to be flipped in the end.

This doesn’t seem like a big issue at first, but we have had so many people confused with flip_y in a competition on Kaggle at https://www.kaggle.com/c/instant-gratification.

[akeshavan]

The current description is: “The fraction of samples whose class are randomly exchanged. Larger values introduce noise in the labels and make the classification task harder.”

We are suggesting: “The fraction of samples whose class is assigned randomly. Larger values introduce noise in the labels and make the classification task harder.”

Any other suggestions?

[TomDLT]

Is it adequate to change the description of the variable and keep the variable name flip_y constant, or is changing the variable name to something like random_fraction an option?

I think we would prefer not changing the variable name, and just improve the documentation.