One-hot encoding flips binary labels

[unai-fa]

Hi,

I noticed that the one_hot() function flips labels when applied to binary factors. Specifically, the function uses contrasts() and then drops one of the two columns in the binary case. This results in "0" being encoded as 1 and "1" as 0, which may be unexpected.

labels <- factor(c(0, 1, 1, 0))
one_hot(labels)
# Output: [1] 1 0 0 1
# Expected: [1] 0 1 1 0

This can lead to inconsistencies if the base model is trained on factor labels and the auditor uses one_hot().

Thanks!

[pfistfl]

Thanks, I'll try to get a fix in soon-ish.
Feel free to also create a PR

[pfistfl]

@unai-fa
I had a look. While I agree that this can lead to confusion, I'm not sure what the correct intended behaviour is.
This seems to be an exception where the factor names are integers starting from 0.
Do you disagree with dropping the second column?