Handling NaNs in NMF

[OlenaBugaiova]

Describe the workflow you want to enable

Motivation:
Sparse matrixes are very common in recommender systems problems. And matrix factorization approach is one of the most popular approaches for this task. But recommender systems often have sparse data with a big amount of missing values
Problem:
In principle, non-negative matrix factorization can work with sparse matrices and optimize the solution based only on the present values. In the scikit-learn implementation, the validation doesn't allow the fit_transform method of NMF to accept sparse matrixes with NaN values

Describe your proposed solution

• In this file https://github.com/scikit-learn/scikit-learn/blob/98cf537f5/sklearn/decomposition/_nmf.py
  In the fit_transform method there is code:
  X = self._validate_data(
  X, accept_sparse=("csr", "csc"), dtype=[np.float64, np.float32]
  )
• The parameter accept_sparse with ("csr", "csc") default values will be passed to the check_array method in https://github.com/scikit-learn/scikit-learn/blob/main/sklearn/utils/validation.py so validation accepts a sparse format
• But the check_array method has a parameter force_all_finite with a default value of True. And this parameter doesn't allow check_array to accept NaN values even when the sparse format ("csr", "csc") is allowed
• As a result, the method def _assert_all_finite of the validation.py file throws the following error: ValueError("Input contains NaN")

Can you make it configurable to add force_all_finite='allow_nan' to the _validate_data method in the fit_transform method of NMF?

Describe alternatives you've considered, if relevant

No response

Additional context

No response

[glemaitre]

Letting the matrix might not be sufficient because a sparse matrix needs to be handled with care and might require to use of specific methods of the scipy.sparse.

ping @jeremiedbb that knows better the internal implementation that can now if it would be more or less feasible?

[jeremiedbb]

The issue seems to be about handling nan rather than supporting sparse matrices (which is already the case). Let me rename the issue.

Regarding nan handling, there is this PR #8474. It would be interesting to revive this PR, but we need to take a look at the litterature to check if the PR implements a common approach.

[OlenaBugaiova]

It makes sense, thanks for looking at this request. I agree, it is about handling Missing Values rather than sparse matrixes.
I see, there was work going on in PR #8474

[TomDLT]

I revived the PR adding support to missing values in NMF (#8474). Note that for now, the PR only adds support to missing values in dense datasets. Tell me if you have any question.

[OlenaBugaiova]

That's wonderful! I am trying it locally.
You added a test that the NMF imputation is better than SimpleImputer+NMF. That's why I was looking for it.

Please let me know if I can help with writing more tests or documentation.

[OlenaBugaiova]

Tested this functionality locally on data having a lot of NaNs. It worked perfectly.
Notes:

1. NMF ‘cd’ solver doesn't handle missing values
2. Init method can not be 'nndsvd' / 'nndsvda', 'nndsvdar' in the case of data with missing values

Closing this feature request

[OlenaBugaiova]

It would be nice to add NMF to this page after merging https://scikit-learn.org/stable/modules/impute.html