MultiOutputClassifier does not rely on estimator to provide pairwise tag

[Alex-Xenos]

Describe the bug

I use the MultiOutputClassifier function to make SVC multilabel.

Then, if I use the linear or rbf kernel the cross_validation function works perfectly fine.

However, when I use SVC with precomputed kernel is having an ValueError: Precomputed matrix must be a square matrix.

Steps/Code to Reproduce

import numpy as np
from sklearn.svm import SVC
from sklearn.metrics import pairwise_distances
from sklearn.multioutput import MultiOutputClassifier
from sklearn.model_selection import cross_val_score, cross_validate

svm = SVC(kernel='precomputed', C=100, random_state=42)
multilabel_classifier = MultiOutputClassifier(svm, n_jobs=-1)

X = np.random.rand(1000, 1000)
y = np.random.randint(0, 2, size=(1000, 6))

kernel_eucl = pairwise_distances(X, metric='euclidean')

cross_validate(
    multilabel_classifier, kernel_eucl, y, cv=10, scoring='f1_weighted', n_jobs=-1
)

Expected Results

An weighted f1-score.

Actual Results

ValueError: 
All the 10 fits failed.
It is very likely that your model is misconfigured.
You can try to debug the error by setting error_score='raise'.

Below are more details about the failures:
--------------------------------------------------------------------------------
10 fits failed with the following error:

 File "C:\Users\bscuser\anaconda3\Lib\site-packages\sklearn\svm\_base.py", line 217, in fit
    raise ValueError(
ValueError: Precomputed matrix must be a square matrix. Input is a 900x1000 matrix.

Versions

System:
    python: 3.11.7 | packaged by Anaconda, Inc. | (main, Dec 15 2023, 18:05:47) [MSC v.1916 64 bit (AMD64)]
executable: C:\Users\bscuser\anaconda3\python.exe
   machine: Windows-10-10.0.19045-SP0

Python dependencies:
      sklearn: 1.2.2
          pip: 23.3.1
   setuptools: 68.2.2
        numpy: 1.26.4
        scipy: 1.11.4
       Cython: None
       pandas: 2.1.4
   matplotlib: 3.8.0
       joblib: 1.2.0
threadpoolctl: 2.2.0

Built with OpenMP: True

threadpoolctl info:
       filepath: C:\Users\bscuser\anaconda3\Library\bin\mkl_rt.2.dll
         prefix: mkl_rt
       user_api: blas
   internal_api: mkl
        version: 2023.1-Product
    num_threads: 4
threading_layer: intel

       filepath: C:\Users\bscuser\anaconda3\vcomp140.dll
         prefix: vcomp
       user_api: openmp
   internal_api: openmp
        version: None
    num_threads: 4

       filepath: C:\Users\bscuser\anaconda3\Library\bin\libiomp5md.dll
         prefix: libiomp
       user_api: openmp
   internal_api: openmp
        version: None
    num_threads: 4

[glemaitre]

It seems that the issue boils to down to not setting properly the pairwise tag for the meta-estimator. So we could solve the issue by including the following in the dictionary returned by _more_tags:

"pairwise": _safe_tags(self.estimator, key="pairwise"),

It delegates the pairwise feature to the underlying estimator. I assume we have the same bug for the MultiOutputRegressor and maybe other meta-estimators in this module.

@Alex-Xenos Do you wish to solve this bug and make a pull-request including the fix and the non-regression tests?

[glemaitre]

Actually, we should probably come with a common test to be sure that all meta-estimators are relying on the underlying estimator to set this tag.

[glemaitre]

Please don't provide an LLM suggestion that does not answer to the question. Here we need to work a common test and not just a non-regression specifically for the estimator.