The docstring is just a bit off: what it means is have the same shape[0] or length.
It converts 1d to 2d. But we don't check about ndim >2 and we should.

What would be the benefit of having several functions?

It's not really necessary to separate things into separate functions, but I was thinking that one might want to check only certain subsets of (1,2,3, and 4). For example, some functions might not care if you send in a 1D or a 2D array, but they might still require that the first dimension be the same in both inputs. Another reason might be readability / code reuse. Again, not really necessary.

As you said, I think the main point is to check for ndim.

+1 for breaking check_arrays up. I've been avoiding that function because I found its use complicated, esp. in estimators accepting sparse matrices. E.g. in LinearClassifierMixin I used atleast2d_or_csr and a custom check for n_features. (That's bad, because it produces a different error message than other estimators.)

why did you find the use complicated?

It has a lot of options and the code is quite complex. It also doesn't do all the input validation that is required at predict time (n_features checking), so I use simpler function there instead. atleast2d_or_csr does what its name suggests.

.. but doesn't check for finite numbers and always enforces the same sparsity type...

I think if we will split up the function, that will lead to code duplication and will make it harder to have consistent checks. Feel free to try, though ;)

Hi I am new and trying to make a contribution. I would like to work on a PR for this issue, I think since the situations where ndim > 2 should be consistent across all estimators it should be the case that the check_arrays function raises an error along the lines of "max array dimension has been exceeded". Guidance will be much appreciated.

Hi @hamsal,

Indeed, you'd want to insert a check for ndim > 2 and raise a ValueError in check_arrays. Then put an appropriate unit test in sklearn/utils/tests/test_validation.py (or adapt an existing test) and run the full test suite on your local clone to see if anything fails and send a PR if it doesn't (or even if it does, but you got close). A PR would be much appreciated!

@larsmans Thanks for the note to include a unit test.

Closing as solved by #2958.