tests are failing

Looked into it. One of the cases where it's failing is in test_gradient_boosting_loss_functions.test_weighted_percentile_zero_weight

The issue is the 100. / weight_cdf[-1] in

p_n = 100. / weight_cdf[-1] * ( weight_cdf - sample_weight / 2. )

when the weights are 0. What behavior is expected in this case? Uniform weighting?

The other case where tests are failing involves test_gradient_boosting.test_boston.

The issue seems to be the interpolation is causing the result to be slightly different than the expected. This seems consistent with introducing interpolation -- e.g., consider the original example:

weight = np.array([1, 1])
data = np.array([0, 1])
_weighted_percentile(data, weight, 50)

Originally, this returned 0. Now it returns 0.5.

It seems like currently in the zero-weight case, the smallest value is chosen, i.e., if percentile_idx = 0 (which happens in the 0-weight case), then sorted_idx[percentile_idx] is simply choosing the min value in the array. (Relevant lines 56-59)

@amueller Sorry for the 2 year delay! Should be ready to merge.

Arfff. It is only now that I am finding this PR. I am sorry I opened #17768 instead.
I will add you in the contributor.

@glemaitre No worries at all, and thanks!

I'll close this PR in favor of yours.