Add ArcSinhTransformer for inverse hyperbolic sine transformation #879
Conversation
ankitlade12
commented
Jan 8, 2026
ankitlade12
commented
- Add ArcSinhTransformer class with loc and scale parameters
- Support for positive and negative values (unlike LogTransformer)
- Includes inverse_transform method
- Add comprehensive tests with pytest parametrize
- Add user guide documentation
There was a problem hiding this comment.
Hi @ankitlade12
This transformer is ready to go (almost ;) )
We need a few additions:
- please add the transformer in docs/index
- please add the transformer to docs/api
- please add to readme
- please add to https://github.com/feature-engine/feature_engine/blob/main/tests/test_transformation/test_check_estimator_transformers.py
- please rebase main onto this branch so that non-related tests pass
- did you check that the docs pass?
In the user-guide, maybe we can say a bit more about how this transformer compares to log and arcsin? so that the user understands when to use this and when to use the others?
It might help to actually have a demo comparing the 3 transformers, highlighting when this one works and the others don't and viceversa.
It would also be helpful to understand why we should scale the data with loc and scale, not just the effect it has on the values, but an intuitive explanation to someone that is less familiar with transformations and wants to understand if they should apply this to their data or not.
Could we add at the end of the user api file the links with more info suggested in the issue:
https://stats.stackexchange.com/questions/1444/how-should-i-transform-non-negative-data-including-zeros
And maybe also a link to the article:
https://www.jstor.org/stable/2288929
- Add ArcSinhTransformer class with loc and scale parameters - Support for positive and negative values (unlike LogTransformer) - Includes inverse_transform method - Add comprehensive tests with pytest parametrize - Add user guide documentation
…ide with comparison and references
ankitlade12
force-pushed
the
add-arcsinh-transformer
branch
from
4037b58 to
201b032
Compare
solegalli
left a comment
solegalli
left a comment
There was a problem hiding this comment.
This transformer is technically good to go. However, the user guide needs a bit of work.
I went over the suggested articles, and my understanding is that this transformation is a good alternative when we have a variable with a distribution that would benefit from log transform, and also a lot of 0s as values, which the log can't handle. We need to discuss that much more.
We need a suitable example (I think we can find one among the datasets from imblearn), and then transform the variable and show the resulting distribution.
In the example shown in this user guide, we achieve the opposite of what we want with a variance stabilizing transformation: we had a more or less normal distribution, and after the transformation, it looks like there 2 distributions. I don't think that is what we want. That result, in fact, is probably one of the limitations of this transformation and a reason why it should not be implemented blindly.
I'll try to give it a go at updating the user docs tomorrow. If you want to go first, by all means do, and I'll pick up from where you left off :)
2 participants
