✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: 7652bde. Link to the linter CI: here}

Thanks for the review @ogrisel! I'm somehow busy rn but will try resolving them ASAP.

@ogrisel Conversations resolved. As for backward compatibility, I added a new keyword sparse_format which is by default None, returning the dense numpy ndarray. Currently I allowed all the existing sparse formats. Since the format may change during computations and supports different operations, I only did .asformat as the very end. Please let me know if this should be improved.

I think we should use a sparse array instead of a sparse matrix. It means that to use this function with the non default sparse_format param, the user would need scipy 1.8 or later but i think this is fine for a new feature.

I think we should use a sparse array instead of a sparse matrix. It means that to use this function with the non default sparse_format param, the user would need scipy 1.8 or later but i think this is fine for a new feature.

Actually, looking at the code again, I think this PR is good the way it is. Let's wait and see at which speed scipy will deprecate scipy sparse matrices in favor of scipy sparse arrays.

In particular, I expect that at some point top level functions used in this PR such as scipy.sparse.eye and scipy.sparse.random will be updated to return sparse arrays instead of sparse matrices and we will have nothing to do in scikit-learn.