Sample weights

[juAlberge]

Hello,
Would it be possible to define a version of the GLEstimator for a proper scoring rule that accept sample weights?
That would be useful for survival analysis to adjust for censoring

[aperezlebel]

Thank you for the suggestion! So far the implemented GLEstimator only supports the Brier score. I'll work on adding an option to specify the scoring rule. This will be straightforward from the paper, except for the bias estimation, which was derived for the Brier score only.

What is the mathematical definition of the weighted proper scoring rule you are considering?

[ogrisel]

weighted_BS = np.mean(
    sample_weight * (model.predict_proba(X_test) - np.eye(n_classes)[y_test]) ** 2
)