"Supervised" models with no predict method

[ablaom]

On this thread @darrencl asks how one should implement a feature selection "transformer" that uses a "target". His use case is feature selection based on strong correlation with the target.

One could implement such a model as Supervised with a transform method; this is allowed. (One could not use predict to output transformations, as the API assumes the output of predict has the same scitype as the target y passed to fit (or is a probablistic version of same). However, the method then has no predict method, which would be unexpected, although possibly not impossible to live with.

An alternative that I recommend is to implement the model as Unsupervised with the "target" being bundled with the other features X by the model user. The model has a hyper-parameter specifiying which feature is the comparison_feature (aka "target"). Then there is no need for a predict method. The only drawback is that the "target" feature itself needs to be added to the input, which would mean adding an element to the overall pipeline. But I don't think that's a big deal. Whether the comparison_feature is itself retained as part of the output could be specified by a second hyperparameter drop_comparison_variable or whatever.

Are there any other suggestions?

[ablaom]

I think my second suggestion above deals with the question posed, and in the absence of other input, I am now closing. Feel free to continue the discussion here or re-open.