[MRG] FIX Revert the addition of ndcg_score and dcg_score #9932
Conversation
This reverts commit 56a21ea.
|
My only qualm with this is that dcg_score might actually work. But who'd know, as it's not tested nor correct behaviour documented. |
I think dcg_score is correct, although the docstring isn't. also, if y_true is binary, I don't see what achieves |
doc/whats_new.rst
Outdated
| @@ -20,3 +20,5224 @@ Previous Releases | |||
| Version 0.14 <whats_new/v0.14.rst> | |||
| Version 0.13 <whats_new/v0.13.rst> | |||
| Older Versions <whats_new/older_versions.rst> | |||
|
|
|||
|
@TomDLT, opinion on the reversion? |
sklearn/metrics/ranking.py
Outdated
| from ..utils.multiclass import type_of_target | ||
| from ..utils.extmath import stable_cumsum | ||
| from ..utils.sparsefuncs import count_nonzero | ||
| from ..exceptions import UndefinedMetricWarning | ||
| from ..preprocessing import label_binarize |
There was a problem hiding this comment.
Seems that we need label_binarize here, it's now used in roc_auc_score for supporting binary y_true other than {0, 1} / {-1, 1}.
|
I agree it's simpler to revert it. We probably need to document the reversion, adding for example a descriptive error for users that will try to import these metrics. |
|
Nah. I don't think we'll get complaints about this ImportError. No one can
have used ndcg_score. Better off not having the name in the module __dict__.
|
|
if we do not provide a message for the revert, do we need to remove the what's new entry in 0.19? |
|
we will note the removal in what's new for 0.19.1.
…On 16 Oct 2017 8:33 pm, "Hanmin Qin" ***@***.***> wrote:
if we do not provide a message for the revert, do we need to remove the
what's new entry in 0.19?
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
<#9932 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AAEz6-TMN7fK7UYh39rFZxwjndArEck5ks5ssyLrgaJpZM4P6E9G>
.
|
|
Actually, we have sufficient consensus. Merging. |
Can you explain that? I'll go ahead with making release preparations, but this smells funny to me. I haven't kept track of the issues as you have, though... |
|
tbh dcg_score could have possibly been used. ndcg only worked if the score
matrix was square, pretty useless in practice. Also it handled a weird
definition of ndcg that was neither normalised nor discounted.
|
This is an alternative to trying to fix up an unusable, undocumented implementation...
Closes #9921
Closes #9928
Closes #9929
Closes #9930
Closes #9931