shouldn't the inverse transform support sparse matrices since the transform method can return a sparse matrix ?

transform method can return a sparse matrix

How to trigger this behaviour? I was not able to.

When fitting on dense X and trying to use inverse_transform on sparse, I got the following error:

encoder.inverse_transform(sparse.csr_matrix(X_tr))
---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
/tmp/xxx.py in 
----> 32 encoder.inverse_transform(sparse.csr_matrix(X_tr))

~/Documents/packages/scikit-learn/sklearn/preprocessing/_encoders.py in inverse_transform(self, X)
    882                 found_unknown[i] = unknown_labels
    883             else:
--> 884                 X_tr[:, i] = self.categories_[i][labels]
    885 
    886         # insert None values for unknown values

IndexError: only integers, slices (`:`), ellipsis (`...`), numpy.newaxis (`None`) and integer or boolean arrays are valid indices

but it might be because I am not able to trigger a sparse X_trans when calling transform.

And I assume that OrdinalEncoder in general should not really benefit from sparse matrix. You would need to always have the category 0 but creating a sparse matrix would be rather weird regarding the semantic of the category).

transform can't return a sparse matrix indeed. The docstring is wrong.
Edit: actually it's not wrong anymore. I was looking at an old version of the code :)

Edit: actually it's not wrong anymore. I was looking at an old version of the code :)

Yes, I might solve it in the previous PR :)