[BUG] OutlierRemover doesn't work with supervised learning

[MBrouns]

X and y will have a different size after the outlier removal because we can't filter y in the pipeline

[Matgrb]

I suppose this could be fixed using imbalanced-learn Pipeline or implementing own pipeline that allows this kind of actions.
What are your thoughts on it?

[koaning]

I like the idea and it's certainly valid statement, but we'd like this library to remain compatible with scikit-learn primarily. I can imagine that it gets complicated for users too if they need to figure out which components require which pipeline backend.

[MBrouns]

I've given it some thought and I do think that a custom pipeline will be the only way to properly support this. That said, there might be something to say for reusing the imblearn pipeline here. If I recall correctly they use a resample method rather than a transform. That allows us to raise a warning in the transform of the outlier remover stating that if its used with supervised learning its probably better to use the imblearn pipelines.

@Matgrb were you thinking about picking this issue up?

[sephib]

Hi @MBrouns,
Thx for all the work on the package.
Until this issue is resolved, what is your strategy to implement an outlier dedaction within a pipeline? just to preform this step in the pre-process step and not as part of the splitting of the data within the pipeline? (I've shamelessly added a link to my blog post related to splitting of the data with dataclass - is this a relevant feature for scikit-lego?

[MBrouns]

Hm I kind of like being able to do as much as possible inside a pipeline. It makes persistence much simpler and there's less chance stuff will go wrong with data splits. I'll take a look at your approach though and see whether that fits!

[sephib]

I kind of like being able to do as much as possible inside a pipeline. It makes persistence much simpler and there's less chance stuff will go wrong with data split

Totally agree - that's why I was wondering what is your current solutions for outliers