RFC Behavior of OneHotEncoder on binary features

Right now OneHotEncoder expands every bindary feature into two features afaik. I’m not sure if this is a great/convenient behavior. I think it might be nicer if (optionally?) it’d use a single column - that’s particularly natural if the feature was already 0 and 1 Otherwise that either makes the ColumnTransformer people have to use more complicated, or creates redundant features. One might argue that one possible cure for that is the option to drop one of the indicator variables. I’m not really sure if that’s what I want, though. In my mind having a base category is more interpretable in the binary case than in the multinomial case.

[justmarkham]

I realize this issue is closed, but I wanted to offer my opinion here in case it’s useful for future discussions of OneHotEncoder.

I saw this comment above from @jnothman:

I’d be happy to add drop=‘binary’ and eventually change the default. But I’m okay with having drop=‘first’ merged as a first go.

Personally, I would advocate for drop=None to continue to be the default. Here’s why:

1. drop=None is consistent and easy to understand. For example, if you have 1 possible value in the feature, OHE creates 1 column. 2 possible values creates 2 columns. 3 possible values creates 3 columns. (And so on.) In contrast, drop='if_binary' is less consistent and thus harder to understand. If you have 1 possible value in the feature, OHE creates 1 column. 2 possible values creates 1 column. 3 possible values creates 3 columns.
2. drop=None is compatible with handle_unknown='ignore', which is a very useful option, whereas drop='if_binary' is incompatible with that option.
3. drop=None is consistent with the get_dummies() function in pandas.

Thus for the majority of cases in which you don’t need to remove redundant columns, drop=None is a sensible default.

For the minority of cases in which you do need to remove redundant columns, drop='first' is already a great option.

Thus for me, drop='if_binary' would be the least desirable default.

[drewmjohnston]

@pavopax there are a couple of other threads in which people discuss some useful information about this–dropping a value seems to matter most in regression (both unregularized and regularized). In the second case, it effects the value of the coefficients, which is important for interpretation. In the second case, this can even effect which columns drop out, which can lead to different solutions. Perfect collinearity also plays poorly with some other models, such as Keras neural networks.