scikit-learn compatibility

[martinfleis]

While I am trying to make our classes compatible with sklearn as much as possible (in #10 we start subclassing BaseEstimator), there are some limits to this at the moment. Notably, the signature of fit in sklearn estimator should be fit(X, y), while we have fit(X, y, geometry). I believe that what we have is the best option but it makes non-conformant.

We fail sklearn.utils.estimator_checks.check_estimator checks due to expected geometry and stuff like GridSearchCV does not work either as it is unable to pass geometry to predict. Same issue applies to score.

We could potentially pass geometry within init but then any CV splits would break anyway. Alternatively, geometry could be part of X but then none of the transformers in sklearn pipelines will work.

I am not sure if there is a way to make our API fully sklearn compatible at the moment.

[ogencoglu]

Is there any tutorial or example in docs on how to use this library and for what?

[martinfleis]

It is very much alpha state of development but some notebooks are in docs -> https://github.com/pysal/gwlearn/tree/main/docs/source.

[theralavineela]

I've been researching Metadata Routing (introduced in sklearn 1.3). It seems like a promising way to handle geometry as routed metadata. This would allow us to keep the fit(X, y, geometry) signature while allowing the model to work inside standard Pipelines.

I’d love to help address this if it's a priority.

[martinfleis]

That would be great!

[theralavineela]

hi @martinfleis , I worked on the sklearn compatibility issue in gwlearn. The main problem was that geometry breaks Pipeline and GridSearchCV.

I enabled sklearn’s metadata routing and updated the base estimator so geometry is routed instead of being a required fit argument. I also added a thin sklearn adapter to safely wrap GW models so they can be cloned and tuned with GridSearchCV.

I verified that geometry is successfully passed through CV and pipelines. The remaining error I hit is just a parameter-naming mismatch in the demo, not a routing issue.

I’d love your thoughts on whether submitting this as a small PR (routing + adapter) makes sense, or if you’d prefer a different direction.

[martinfleis]

PR please :)

[theralavineela]

Sure, I’ll open a PR shortly.

[yousefkhedr777]

Hi @martinfleis ,I updated the BaseRegressor to support metadata routing for geometry . i added it to fit ,predict ,score methods so it works good with GridSearch and pipeline and kept the signatures clean using keyword only arguments .