using compute_sample_weight for the fit_params of cross_val_predict

[AlexandraBomane]

I would like to know if it is possible to use "compute_sample_weight" within "cross_val_predict" to provide sample weights in fit_params?

I'm using the following code:
y_pred = cross_val_predict(clf, X, y, cv = cv, method = 'predict', fit_params = {'sample_weight':compute_sample_weight("balanced", y)})

But I think that it is not correct because compute_sample_weight("balanced", y) computes weights for all labels, whereas I would like that this function successively applied on each training set generated by the split operated by cross_val_predict. I think that the parameter "indices" of compute_sample_weight() could be a solution, but I don't know how to apply it.

[jnothman]

No there is no straightforward way to do this at the moment. Consider using the Pipeline from the imblearn library for now. This will be solved with #3855 I hope.

[AlexandraBomane]

Which solution from this post could solve my problem?

[jnothman]

Use Imbalanced learn: https://imbalanced-learn.readthedocs.io/en/stable/api.html

[AlexandraBomane]

I meant which feature of Imbalanced learn could help? Because I don't want to resample X and y, I just want to recalculate sample_weight according to X and y.

[jnothman]

Yes, I can't recall whether the last release allowed the resampler to modify a changed sample_weight. We hope to support that case here eventually...