KBinsDiscretizer produces wrong bins with repeated small values

The KBinsDiscretizer will fail to produce quantile discretizations if the input data has most of its entries having the same value and corresponding to the first bin:

import numpy as np
from sklearn.preprocessing import KBinsDiscretizer
kb = KBinsDiscretizer(n_bins=3, encode="ordinal", strategy="quantile")
a = np.array([0,0,0,0,0,0,0,0,1,2]).reshape((-1,1))
kb.fit(a).transform(a)

/home/david/anaconda3/lib/python3.7/site-packages/sklearn/preprocessing/_discretization.py:222: UserWarning: Bins whose width are too small (i.e., <= 1e-8) in feature 0 are removed. Consider decreasing the number of bins.
  'decreasing the number of bins.' % jj)

array([[0.],
       [0.],
       [0.],
       [0.],
       [0.],
       [0.],
       [0.],
       [0.],
       [0.],
       [0.]])

In this case, there’s no reason why the last two elements could not be assigned to the highest bucket instead of being grouped with the rest.

[david-cortes]

But usually the reason why one wants to use the quantile strategy is because the distribution to be transformed is skewed or multi-modal or something like that, for which uniform is not useful.

Also in that example the right output (assuming quantiles) would be to assign the last two observations to the same bucket, since values greater than zero constitute less than a third of the input.

[azihna]

I just wanted to clear one wrong thing I said up there: during transform, the bin edges are actually replaced with -inf, inf. As seen here. Sorry for any confusion about that.