This is intended to be a quick fix for 0.19.1. I am creating other issues to address shortfalls in ndcg API and testing.

Sorry to disturb if I'm not qualified to post my opinion here.
I don't think the PR fixes #9921, because seems that we still can't run the following script posted at the beginning of the issue:

y_true = [0, 1, 0, 1]
y_score = [[0.15, 0.85], [0.7, 0.3], [0.06, 0.94], [0.7, 0.3]]
metrics.ndcg_score(y_true, y_score)

There seems two problems and the second is not solved here:
(1)The shape of binarized_y_true is not correct. It seems that this can be solved by using y_score.shape[1]
(2)For LabeBinarizer, binary targets is transformed to a column vector, which is not we want. An alternative solution might be using OneHotEncoder.

Also, if current implementation is right, could we provide a reference for users (and for me :) )? I can't find any reference which is consistent with current implementation and the reference in the doc is a dead link. Personally, I might still don't think current implementation is right. It's a simple copy-paste from kaggle and I think ogrisel's implementation here is at least what I would use.