multiclass - type_of_target returning 'multiclass-output" despite size being 1? #19629

Isn't this a 1 x 3 array. Both their dimension do not satisfy size > 1 (column does row doesn't)

It's still a 2D array, meaning a numpy with 2 axis:

>>> y = np.array([[1,2,3]])
>>> y.ndim
2
>>> y.shape
(1, 3)

The first axis has size 1 and is expected to represent the number of samples. The second axis has size 3 and is expected to represent number of targets.

If you want to do traditional (single target) multiclass, please just pass the column you want directly as a 1D array. For instance y = np.array([1,2,3]) whould have shape (3,) and would be interpreted as an array of class labels for 3 samples.

You're right. np.array([[1, 2, 3]]) has shape (1, 3), so its first dimension is not greater than 1.

The current documentation says that multiclass-multioutput requires both dimensions to be greater than 1, but:

type_of_target(np.array([[1, 2, 3]]))

returns:

'multiclass-multioutput'

This is because the implementation classifies a 2D array with more than two unique discrete values as multiclass-multioutput before the (n_samples > 1 and n_outputs > 1) distinction implied by the documentation.

So the issue is essentially a documentation mismatch/ambiguity, rather than something wrong with your interpretation of the shape.

For comparison:

np.array([[1, 2, 3]]).shape
# (1, 3)

type_of_target(np.array([[1, 2, 3]]))
# 'multiclass-multioutput'

The documentation should therefore be clarified to accurately describe this edge case.